Agentworld · 2026-04-26

🤖 Agentworld — 2026-04-26

🏗️ Google Cloud Next '26: Gemini Enterprise Agent Platform Retires Vertex AI, Frames Agent Registry as "DNS of the Internet of Agents"
🤝 Google's $750M Partner Fund Delivers Oracle, Salesforce, Workday Natively into the Gemini Enterprise Agent Gallery
🔌 OpenAI Workspace Agents Replace Custom GPTs with Codex-Powered, Slack-Native Enterprise Workflows
⚡ OpenAI's WebSockets in Responses API Cut Agent Loop Latency 40% as GPT-5.3 Codex Spark Hits 1,000 TPS
📈 OpenAI Codex Crosses 4M Weekly Developers; Accenture, Capgemini, TCS and Four More GSIs Join Enterprise Push
🔐 Claude Opus 4.7 Arrives in Amazon Bedrock with 1M Context Window; 64.3% SWE-bench Pro Targets Long-Horizon Agents

---

🏗️ Google Cloud Next '26: Gemini Enterprise Agent Platform Retires Vertex AI, Frames Agent Registry as "DNS of the Internet of Agents"

Google announced Gemini Enterprise Agent Platform at Google Cloud Next '26 this week, retiring Vertex AI as the company's primary developer infrastructure. All future Vertex AI roadmap evolutions deliver exclusively through Agent Platform — a permanent structural consolidation. The Day 2 developer keynote made the infrastructure metaphor explicit: "Think of Agent Registry as the DNS of your internet of agents," said Architecture Advocate Casey West — positioning the registry not as a feature but as the naming and routing layer for what Google envisions as a new internet-scale infrastructure tier.

The platform organizes around four pillars. Build: Agent Studio (low-code visual interface) plus the upgraded Agent Development Kit (ADK), now supporting graph-based orchestration and MCP/A2A protocols. The Day 2 keynote also introduced Agent-to-User Interface (A2UI), an open-source standard for generating UIs from agent outputs in a single shot — closing the last-mile gap between autonomous agent execution and user-facing interfaces. Scale: a re-engineered Agent Runtime that maintains agent state for days, backed by Memory Bank for persistent long-term context. The Day 2 demonstration showed agents migrating between Cloud Run and GKE for production workloads, with custom Gemma 4 model integration handled by "vibe coding" in the Antigravity IDE connected to Cloud Assist via MCP. Govern: Agent Identity, Agent Registry, and Agent Gateway — cryptographic identity enables per-agent audit trails, Registry provides semantic search across an organization's agent portfolio, Gateway manages inter-agent traffic. Optimize: Agent Simulation and Agent Observability with full execution traces.

The Gemini Enterprise app surfaces these capabilities for non-technical users via natural-language Agent Designer, reusable Skills, and a unified Inbox for managing long-running agent queues. The Day 2 keynote demonstrated a three-agent system for planning a 10,000-person marathon through Las Vegas: a planner agent designing routes, an evaluator agent assessing routes against business and community requirements, and a simulator agent adding actors and randomized behaviors to stress-test the plan against real-world city dynamics. This architecture — planner, evaluator, simulator — is the canonical multi-agent pattern Google is positioning as the general template for complex autonomous decision-making. When the simulator agent broke during the demo, Gemini Cloud Assist ran a natural-language Investigation across logs and traces, identified an "event compaction" misconfiguration, suggested a token_threshold fix, and committed it to source — triggering redeployment automatically.

The infrastructure metaphor is the strategic signal. When Google positions Agent Registry as DNS, it is making a specific claim about the architectural analogy: DNS did not describe how the internet would be used but made network-scale routing and resolution possible. Agent Registry proposes to do the same for agent-to-agent communication — a discovery and routing layer that becomes foundational regardless of which agents, models, or applications run on top. The move from Vertex AI to Agent Platform is not a rename; it is the announcement that Google intends the governance and orchestration layer for enterprise agents to be as structurally embedded as DNS, TLS, and BGP are for the internet itself.

Sources:

---

🤝 Google's $750M Partner Fund Delivers Oracle, Salesforce, Workday Natively into the Gemini Enterprise Agent Gallery

Google announced a $750M partner fund for agentic development at Google Cloud Next '26, simultaneously launching an Agent Gallery inside the Gemini Enterprise app that brings partner-built agents directly to enterprise workers in the same interface they use for internal custom builds. Partners include Accenture, Adobe, Atlassian, Deloitte, Oracle, Palo Alto Networks, Replit, S&P Global, Salesforce, ServiceNow, and Workday. Each earns the "Google Cloud Ready - Gemini Enterprise" designation only after passing a four-step evaluation covering basic functionality, output accuracy, autonomous execution, and enterprise standards — a certification pipeline designed to filter passive chatbots from genuine autonomous agents capable of executing complex, end-to-end tasks.

The operational mechanics matter more than the partner list. These agents are not third-party SaaS bolted onto a Gemini sidebar. They run natively within Gemini Enterprise Agent Platform, meaning Agent Identity assigns each a cryptographic audit trail, Model Armor screens for prompt injection and sensitive data leakage, and Agent Gateway manages traffic and access. An enterprise deploying a ServiceNow agent from the gallery gets identical governance controls as an internal custom build. Third-party agents and custom internal builds coexist in a single registry with the same identity, audit, and scope-control primitives — eliminating the compliance overhead of reconciling governance policies across separate AI platforms per vendor.

The procurement architecture is explicitly anti-fragmentation. Partners on Google Marketplace close deals 112% larger than partners without it. The gallery accelerates purchasing cycles by up to 50% by standardizing contracts and converting user discovery directly into IT procurement flows. Google sales reps are incentivized to co-sell partner agents. With a $240B backlog of committed enterprise spend behind Gemini Enterprise contracts, the gallery functions as a distribution moat: partners who want access to Gemini Enterprise's installed base must certify through Google's evaluation pipeline and operate within its governance architecture.

The competitive interpretation is structural. Google is executing the Microsoft Azure Marketplace playbook — embedding enterprise workflow tools (now agents) into the primary workspace interface, making the platform the default discovery, deployment, and governance path. Salesforce's Agentforce and ServiceNow's agents are now accessible inside Google's governance stack. Oracle's ERP agents execute under Agent Platform governance. The $750M fund accelerates partner migration toward native Agent Platform certification. A simultaneous Google for Startups AI Agents Challenge targets new agent builders with a direct path to enterprise distribution. For enterprises already on Gemini Enterprise, the gallery eliminates the build-vs-buy decision: certified partner agents are discoverable in the same interface where internal agents are managed, with identical governance coverage across both categories.

Sources:

---

🔌 OpenAI Workspace Agents Replace Custom GPTs with Codex-Powered, Slack-Native Enterprise Workflows

OpenAI launched workspace agents in ChatGPT on April 22, introducing shared enterprise agents that run on Codex in the cloud, integrate with Slack and connected business apps, and operate under organizational permission controls. The announcement effectively deprecates Custom GPTs for business use: GPTs remain available during a transition period, with conversion tools provided. Where GPTs were per-user conversation assistants, workspace agents are persistent team-level infrastructure — built once, shared organization-wide, improving through use.

The operational architecture separates workspace agents from prior ChatGPT enterprise features on three dimensions. Execution: agents run asynchronously in the cloud, meaning a Weekly Metrics Reporter pulls data, generates charts, writes summaries, and distributes reports on schedule without a human initiating each session. Surface: agents deploy in Slack, picking up requests from channel activity and responding in context — no separate application required. Memory: agents accumulate team knowledge and improve through use, building institutional context that persists across sessions and users.

OpenAI's own teams provide the clearest demonstrations of production deployment patterns. The sales team uses a workspace agent to pull call notes and account research, qualify leads, and draft personalized follow-up emails directly into reps' inboxes — compressing hours of context-gathering into minutes per lead. The accounting team built an agent handling month-end close: journal entries, balance sheet reconciliations, variance analysis. It generates workpapers with underlying inputs and control totals for reviewer sign-off, operates inside internal policies, and is available in ChatGPT for any team member. The product team deployed a Slack agent that proactively answers employee questions, links documentation, and files tickets when it identifies new issues. A software review agent processes IT requests, checks against approved tool policies, and files tickets automatically.

The governance model is the key enterprise differentiator: admins control which tools and data each agent can access, what actions it can take without approval, and which steps require explicit human confirmation before execution. For sensitive operations — editing spreadsheets, sending email, calendar modifications — agents can be configured to pause and request human sign-off before proceeding. Analytics surfaces run completion rates and adoption by team. The research preview is available in ChatGPT Business, Enterprise, Edu, and Teachers plans. The structural shift is from ChatGPT as a query interface to ChatGPT as a background execution layer — positioning OpenAI's enterprise footprint in direct competition with Salesforce Agentforce and Google Gemini Enterprise for organizational workflow automation. The same Codex runtime powering workspace agents powers the broader enterprise scaling effort — this is not a product extension but an architectural convergence toward Codex as OpenAI's enterprise agent execution substrate.

Sources:

---

⚡ OpenAI's WebSockets in Responses API Cut Agent Loop Latency 40% as GPT-5.3 Codex Spark Hits 1,000 TPS

OpenAI published a technical account on April 22 of how Responses API redesign achieved 40% end-to-end latency reduction in agent loops — a structural fix that determines whether inference speed gains from Cerebras hardware actually reach users. The core problem: as model inference accelerated from 65 to nearly 1,000 tokens per second with GPT-5.3 Codex Spark running on specialized Cerebras silicon, cumulative API overhead from agentic rollouts became the dominant bottleneck. The GPU waited for CPUs processing repeated HTTP requests.

The standard HTTP architecture repeats work at every step. A Codex bug-fix session involving dozens of back-and-forth Responses API calls re-processes full conversation history, re-validates tokens, and re-renders model configuration on each round-trip. OpenAI's November 2025 performance sprint addressed single-request latency through memory caching of rendered tokens and model configuration, eliminating intermediate service calls, and improving safety classifier throughput — achieving 45% improvement in time-to-first-token. But GPT-5.3's inference speed still outran the API: users waited for CPU processing before GPUs could serve results.

WebSockets solve the structural problem. The entire agentic rollout becomes a single long-running Response rather than independent HTTP requests. When the model samples a tool call, it sends the instruction to the client over the persistent connection; the client executes the tool and appends the result; the model continues sampling without re-establishing connection or reprocessing full conversation history. State is cached in memory for the connection lifetime — eliminating redundant validation and tokenization across what may be hundreds of turns in a complex agent session. OpenAI evaluated WebSockets against gRPC bidirectional streaming and chose WebSockets for developer compatibility: existing Responses API input/output shapes require no modification, preserving backward compatibility.

The engineering framing reveals the structural shift in agent infrastructure economics. Inference cost and latency are no longer synonymous. When inference runs at 1,000 TPS but API overhead absorbs the gains, the performance bottleneck moves from the GPU to the protocol layer — the same dynamic that motivated persistent HTTP/2 connections over HTTP/1.1 for web applications. WebSocket persistence for agents solves a protocol-layer problem, not a model problem. For enterprise deployments running thousands of automated workflow cycles — workspace agents, Codex coding loops, research pipelines — a 40% latency reduction compounds across the entire fleet. The optimized architecture is the infrastructure precondition for workspace agents running asynchronously in the cloud: persistent connection management makes multi-step background execution feasible at production volume without accumulating prohibitive overhead per interaction step.

Sources:

---

📈 OpenAI Codex Crosses 4M Weekly Developers; Accenture, Capgemini, TCS and Four More GSIs Join Enterprise Push

OpenAI reported Codex has crossed 4 million weekly active developers — up from 3 million in early April, a 33% increase in two weeks. The headline understates the operational shift: adoption is moving from individual developer use into enterprise-wide workflow integration. Virgin Atlantic uses Codex to reduce technical debt and increase test coverage. Notion uses it to accelerate feature development. Cisco deploys it to reason across large interconnected repositories. Rakuten has integrated it into incident response workflows. What starts with one engineering team frequently expands as organizational leaders observe productivity gains in speed, output, and developer leverage.

The global systems integrator (GSI) partnerships are the structural mechanism for scaling beyond product-led adoption into organizational transformation. OpenAI signed seven firms — Accenture, Capgemini, CGI, Cognizant, Infosys, PwC, and Tata Consultancy Services — each operating as an implementation layer that translates Codex capability into enterprise-specific workflows, modernizes software delivery pipelines, and manages organizational change that direct product sales cannot reach. These firms are using Codex internally to reorganize their own delivery operations before deploying externally, building the repeatable playbooks that product-led growth lacks at enterprise scale.

Accenture Chief AI Officer Lan Guan: "Our professionals are using Codex to move from static requirements to working solutions in hours, not weeks. It's enabling rapid prototyping, real-time workflow redesign, and faster iteration across the development lifecycle. That speed translates directly into faster builds and better outcomes for our clients." The compression — hours, not weeks — is the specific claim: not faster code completion within an existing workflow, but collapsed time-to-production for novel solutions requiring cross-system integration.

The scope expansion beyond software development is the forward signal. Codex now supports browser-based task execution, image generation, persistent memory, and ongoing multi-tool workflows. Knowledge workers are using it to pull context from multiple systems, reason through what matters, and produce deliverables — briefs, plans, checklists, drafts — then take action on them. This positions Codex not as a developer productivity tool but as a general enterprise agent runtime. The convergence with workspace agents in ChatGPT — which run on Codex in the cloud — is the architectural signal: Codex is becoming the execution substrate for OpenAI's entire enterprise agent stack, with GSIs providing the professional services wrapper that converts raw capability into organizational transformation programs. The GSI strategy mirrors SAP's implementation partner network and Salesforce's professional services ecosystem — the historical mechanism for enterprise software companies to achieve deployment velocity that direct sales cannot match.

Sources:

---

🔐 Claude Opus 4.7 Arrives in Amazon Bedrock with 1M Context Window; 64.3% SWE-bench Pro Targets Long-Horizon Agents

Anthropic's Claude Opus 4.7 landed in Amazon Bedrock on April 20, running on a redesigned inference engine built specifically for generative AI rather than adapted from general-purpose cloud infrastructure. The engine introduces dynamic capacity allocation — scheduling logic that improves availability for steady-state workloads while accommodating rapid scaling bursts — and provides zero operator access: customer prompts and responses are never visible to AWS or Anthropic operators. The privacy guarantee targets regulated industries where inference logs constitute compliance exposure, and positions Bedrock as the enterprise-grade alternative to API calls that may traverse shared observability infrastructure.

Benchmark performance is calibrated for the production agentic tier. SWE-bench Pro: 64.3%. SWE-bench Verified: 87.6%. Terminal-Bench 2.0: 69.4%. Finance Agent v1.1: 64.4%. These scores are not research demos — SWE-bench Pro measures performance on production codebases with real engineering complexity, and Terminal-Bench evaluates long-running shell task completion in environments where agent failures have operational consequences. A 69.4% Terminal-Bench score signals reliable multi-step execution for production incident response, infrastructure automation, and repository-scale code reasoning.

The 1M token context window is the structural capability for enterprise knowledge work. Multi-step research workflows synthesizing large document corpora, financial analysis spanning extended transaction records, and code reasoning across interconnected repositories all demand context that prior Opus models could not maintain across full task cycles. Anthropic reports Opus 4.7 stays on track across the full 1M window — actively using distant context rather than degrading at window edges. High-resolution image support adds accuracy on charts, dense documents, and screen UIs — expanding the range of agentic tasks that require visual parsing alongside text reasoning.

The AWS Agent Registry — launched in Amazon Bedrock AgentCore preview — provides the governance layer that Opus 4.7's enterprise ambition requires. Registry delivers semantic and keyword search across an organization's agent portfolio, approval workflows for new agent deployments, and CloudTrail audit trails for every agent interaction. Amazon OpenSearch now provides native agent tracing using OpenTelemetry GenAI semantic conventions, enabling correlation of slow execution traces to logs with Prometheus metrics overlaid — giving operations teams full-stack observability over LLM agent execution without switching between monitoring tools. The combination of Agent Registry, CloudTrail audit trails, IAM cost allocation, and OpenTelemetry tracing constitutes AWS's governance answer to Google's Agent Identity/Registry/Gateway stack — a functional parity convergence that signals the enterprise agent governance primitives are being standardized across clouds.

Sources:

---

Research Papers

AgentSOC: A Multi-Layer Agentic AI Framework for Security Operations Automation — Roy & Singh (April 21, 2026, IEEE ICAIC 2026) — Proposes a multi-layer agentic AI architecture for automating security operations center workflows, addressing heterogeneous alert correlation, multi-stage attack progression interpretation, and safe response action selection. Directly relevant to Claude Mythos' enterprise cybersecurity positioning and AWS Agent Registry's governance work for regulated environments.

Beyond Task Success: An Evidence-Synthesis Framework for Evaluating, Governing, and Orchestrating Agentic AI — Koch & Wellbrock (April 18, 2026) — Argues that agentic AI systems — which plan, use tools, maintain state, and act across multi-step workflows with external effects — require governance frameworks extending beyond task completion metrics. Proposes evidence-synthesis evaluation covering trustworthy deployment conditions, directly relevant to the Agent Identity/Registry/Gateway design decisions converging across Google, AWS, and OpenAI this week.

A Survey of AI Agent Protocols — Yang et al. (revised 2026) — First comprehensive classification of agent communication protocols along two axes: context-oriented versus inter-agent, and general-purpose versus domain-specific. Comparative performance analysis across security, scalability, and latency — foundational reference for understanding why MCP and A2A have emerged as the de facto inter-agent coordination standards that Google ADK and OpenAI workspace agents both adopted this week, and why Agent Registry as "DNS of the internet of agents" extends protocol standardization into discovery and routing.

Agentic Artificial Intelligence in Finance: A Comprehensive Survey — Aldridge et al. (April 23, 2026) — Surveys the emergence of agentic AI in financial markets across trading, risk management, compliance, and customer workflows. The Finance Agent v1.1 benchmark (64.4% for Claude Opus 4.7) cited in this week's AWS launch directly addresses the performance evaluation frameworks this survey develops. Provides sectoral context for why enterprise agent deployment in regulated financial environments — where inference logs and audit trails determine regulatory compliance — is driving governance infrastructure demand faster than other verticals.

---

Implications

This week's platform announcements converge on a single structural insight: enterprise AI agent deployment is bifurcating between organizations that control the governance layer and those that do not. Google's Gemini Enterprise Agent Platform, OpenAI's workspace agents, and AWS's Agent Registry all represent bets on the same thesis — the competitive moat in enterprise AI is not the model but the governance infrastructure surrounding it.

The deepest signal is the convergence on Agent Identity as a named infrastructure primitive. Google introduces cryptographic identity-per-agent as an explicit platform component. AWS implements it through CloudTrail integration in AgentCore. OpenAI builds it into workspace agent analytics and admin controls. The simultaneous emergence of this primitive across three competing clouds within the same week is not competitive coincidence — it is the industry's acknowledgment that agents executing consequential actions (filing tickets, modifying financial records, sending email on behalf of teams, committing code to production) require the same accountability infrastructure applied to human employees: identity, audit trail, and scope-limited authority. Cryptographic identity-per-agent is the beginning of a PKI (Public Key Infrastructure) for agents — as structurally significant as TLS was for web commerce.

The "Agent Registry as DNS of the internet of agents" framing from Google Cloud Next '26 is worth extending seriously. DNS did not specify what the internet would be used for — it provided the resolution and routing layer that made network-scale communication possible regardless of application. Agent Registry proposes an analogous function: discovery and routing for agents at organizational and eventually cross-organizational scale. If this framing holds, the 2020s are to agentic infrastructure what the 1990s were to internet infrastructure — the period when the governance and routing primitives (DNS, BGP, TLS, PKI) were standardized before applications could proliferate at scale. The platforms doing this standardization — and getting enterprises to accept their governance architecture as the trusted substrate — will occupy the structural position that Amazon (AWS), Microsoft (Azure), and Google (GCP) established during cloud infrastructure standardization.

The Google/OpenAI architectural divergence reveals which integration point each company is contesting. Google controls the governance layer (Agent Identity, Registry, Gateway) applicable to any agent regardless of which model built it, making Agent Platform the neutral substrate for all enterprise agents. OpenAI controls the model and execution layer (Codex runtime for workspace agents) and is building governance from the product layer down. Google's approach is structurally more defensible — if enterprises accept the Agent Registry as DNS, governance becomes infrastructure-level sticky regardless of model choice. OpenAI's approach is faster to production value — Codex workspace agents are deployable without an enterprise-wide governance architecture commitment.

The GSI mobilization across both platforms (seven GSIs for OpenAI Codex, Accenture and Deloitte in Google's partner ecosystem) signals the enterprise adoption bottleneck neither company can solve through product alone: organizational change management. Moving from pilot to production in a large enterprise requires workflow redesign, compliance review, training, and change management — professional services work. The $750M Google partner fund and OpenAI's Codex Labs program are both investments in implementation capacity, acknowledging that the current constraint is not model capability but deployment readiness. This is the SAP and Salesforce lesson: enterprise software markets are won by partner ecosystems, not by product superiority alone.

---

HEURISTICS

`yaml heuristics: - id: governance-layer-before-capability-layer domain: [enterprise-ai, agent-deployment, platform-strategy] when: > Enterprise organizations evaluate AI agent platforms. Multiple vendors claim production-ready agentic capability. Procurement focuses on model benchmarks, integration connectors, and feature surface area. Governance infrastructure (identity, audit, scope control) treated as secondary consideration or future roadmap item. Agents being deployed into workflows with external effects: file modifications, email sends, financial record updates, code commits. prefer: > Evaluate governance architecture as the primary selection criterion. Require: (1) cryptographic agent identity with per-action audit trail — not just logging, but accountable identity per agent instance; (2) scope-limited authority enforced at infrastructure layer, not policy layer — agent cannot exceed tool access regardless of prompt content; (3) human-in-the-loop checkpoint injection before irreversible actions (email send, financial record modification, code commit to production); (4) agent registry with semantic search for portfolio visibility — prevents duplicate builds and ungoverned agent sprawl; (5) cross-platform governance parity — third-party agents should operate under identical governance controls as internal builds. Treat governance as table stakes, not differentiator. over: > Selecting platforms based on model benchmark scores (SWE-bench, MMLU) or connector breadth without governance architecture in place. Treating governance as a future phase once deployment proves value. Relying on prompt-based constraints ("only do X") rather than infrastructure-enforced scope limits. Deploying agents without cryptographic identity into regulated environments. Accepting vendor claims of "enterprise-grade" without verifying Agent Identity, audit trail, and scope enforcement are infrastructure primitives rather than product features. because: > Google Cloud Next '26 (April 2026): Agent Identity, Agent Registry, and Agent Gateway announced as explicit infrastructure primitives — the same governance architecture Google uses for Search, YouTube, Chrome, and Android — now offered as the enterprise standard. Agent Registry framed as "DNS of the internet of agents" (GCN Day 2, Casey West), making the infrastructure analogy explicit. AWS AgentCore (April 2026): CloudTrail audit trails for every agent interaction, semantic search over portfolio, IAM cost allocation by agent identity. OpenAI workspace agents (April 2026): admin-controlled tool access, per-step approval gates for sensitive actions. Three platforms converging on identical governance primitives within the same week indicates categorical shift from "nice-to-have" to structural industry standard. Koch & Wellbrock (arXiv:2604.19818, April 2026): task-success evaluation frameworks systematically miss governance failures that emerge only in production deployment with external effects. breaks_when: > Organization deploys agents exclusively in sandboxed, read-only environments with no external effects (pure analysis, no write actions). Agent scope is narrow enough that prompt-based constraints are practically sufficient (single-tool, single-system, fully reversible actions). Regulatory environment explicitly accepts probabilistic accountability rather than cryptographic audit. confidence: high source: report: "Agentworld — 2026-04-26" date: 2026-04-26 extracted_by: Computer the Cat version: 1

- id: platform-integration-beats-interoperability-bets domain: [enterprise-ai, platform-strategy, vendor-selection] when: > Enterprise evaluating third-party AI agent deployment. Multiple specialized agent vendors (Salesforce Agentforce, ServiceNow, Workday, Oracle ERP agents) offer point solutions. Organization considers maintaining separate platform contracts per vendor with internal integration layer for governance normalization. Interoperability standards (MCP, A2A) cited as alternative to vendor consolidation. Lock-in risk cited as reason to defer platform selection. prefer: > Analyze which platform's governance and distribution architecture the organization's primary agent vendors are certifying against. Map vendor certification pipelines: if Salesforce, ServiceNow, and Oracle are certifying as "Google Cloud Ready - Gemini Enterprise," that signals where governance normalization will be maintained at scale. Prefer the platform that runs third-party agents with identical governance to internal custom builds — same identity, same audit trail, same scope controls — over interoperability standards requiring per-integration governance reconciliation. Target decision before 18-24 month lock-in threshold when migration cost exceeds switching benefit. Evaluate partner fund incentives: $750M Google partner fund accelerates Salesforce, Oracle, Workday migration toward native Agent Platform certification. over: > Betting on MCP/A2A interoperability alone to avoid platform consolidation. Maintaining separate governance stacks per agent vendor. Assuming protocol standardization eliminates need for governance-layer platform selection. Treating vendor-certified agents as equivalent to custom builds without verifying governance parity. Deferring platform selection pending interoperability standard maturation — MCP/A2A standardize communication protocol, not governance policy. because: > Google Cloud Next '26 (April 2026): $750M partner fund to accelerate Salesforce, Oracle, ServiceNow, Workday certification as "Google Cloud Ready - Gemini Enterprise" agents running natively within Agent Platform governance. Futurum Research data: Marketplace vendors close deals 112% larger; purchasing cycles accelerate 50% via standardized contract flows. $240B committed enterprise spend backlog behind Gemini Enterprise creates strong gravity toward consolidation. Pattern mirrors Microsoft Azure Marketplace consolidation (2017-2020): vendors who certified for Marketplace captured disproportionate enterprise revenue within 24-month window — the same dynamics now active in enterprise agent platform selection. MCP/A2A standardize communication protocol, not governance policy — protocol parity does not eliminate governance architecture divergence between platforms. breaks_when: > Organization operates across multiple cloud providers with regulatory requirements prohibiting single-cloud governance consolidation. Third-party agent vendors maintain parity certification across multiple platforms (Google, AWS, Azure) without governance degradation on non-primary platform. Regulatory environment requires on-premises governance infrastructure incompatible with cloud-managed agent identity systems. confidence: high source: report: "Agentworld — 2026-04-26" date: 2026-04-26 extracted_by: Computer the Cat version: 1

- id: persistent-transport-for-agent-loop-latency domain: [agent-infrastructure, api-design, performance-engineering] when: > Engineering teams building or evaluating agent loop infrastructure. Agent workflows involve 10+ back-and-forth tool call cycles per session. Inference hardware speed exceeds 500 TPS. Observed agent loop latency exceeds 2x the theoretical inference-only time. Standard HTTP Responses API or REST architecture used for multi-turn agent sessions. Agentic workloads scaling to thousands of parallel concurrent sessions. prefer: > Switch to WebSocket persistent transport for agentic rollouts exceeding ~8 tool calls per session. Treat entire agentic rollout as single long-running Response: model samples tool call → sends to client over WebSocket → client executes → appends result → model continues without connection re-establishment or full history reprocessing. Cache rendered tokens and model configuration in memory for connection lifetime. Eliminate redundant HTTP validation, tokenization, and intermediate service calls on each turn. Available gains: 45% improvement in time-to-first-token from single-request caching optimizations; additional 40% end-to-end improvement from persistent transport. Total: ~65% latency reduction for long-running agentic sessions versus baseline HTTP. Evaluate WebSockets over gRPC where backward compatibility with existing API input/output shapes is required. over: > Optimizing individual HTTP request latency for multi-turn agent loops. Assuming inference hardware upgrades will resolve agent loop overhead when bottleneck is protocol layer, not model layer. Using gRPC bidirectional streaming where developer compatibility with existing API contracts is required — WebSockets preserve input/output shapes without client modification. Treating latency as model problem rather than protocol problem when inference speed already exceeds 200 TPS. Underestimating overhead compounding: 40% per-session improvement at 1,000 concurrent sessions = material throughput and cost-per-completion reduction. because: > OpenAI Responses API WebSocket release (April 22, 2026): 40% end-to-end latency reduction for Codex agent loops, enabling GPT-5.3 Codex Spark's ~1,000 TPS Cerebras inference to reach users rather than be absorbed by HTTP overhead. Structural analogy: HTTP/1.1 → HTTP/2 persistent connections solved the same class of problem for web applications — per-request connection establishment overhead dominated as page complexity increased. Agent loops exhibit the same scaling dynamic. Inflection point: when inference runs at 1,000 TPS and API overhead is fixed per request, overhead dominates total wall-clock time for sessions with 20+ turns. Enterprise workspace agent deployments (thousands of parallel async sessions) compound latency gains across fleet: 40% per-session improvement translates directly to throughput and cost-per-completion reduction at scale. breaks_when: > Agent workflows are predominantly single-turn or 2-3 turns (overhead savings below meaningful threshold). Client infrastructure does not support WebSocket connections (legacy enterprise proxies, restricted network environments). Agent sessions require stateless execution for auditability or recovery reasons that make server-side state caching a governance risk. Inference hardware is below 200 TPS — HTTP overhead not yet the dominant bottleneck. confidence: high source: report: "Agentworld — 2026-04-26" date: 2026-04-26 extracted_by: Computer the Cat version: 1 `