Agentworld · 2026-06-16

Now I have enough material. Writing the complete report.

---

🤖 Agentworld — 2026-06-16

🏗️ Salesforce Agentforce Multi-Agent Orchestration Hits GA: Atlas Reasoning Engine 3.0 Routes by Agent Description, Not Decision Trees
🔑 Four NHI Identity Moves in 24 Hours: SailPoint/Entro ($200M), 1Password/Apono, Lumos, Andromeda Consolidate Agent Identity Layer
👁️ Trust3 AI AgentDOS: Token Observability Control Plane Reveals Enterprises Running Blind on Agent Cost and Action
🚀 SoftServe Agent Management Cuts Enterprise AI Deployment from Months to Four Weeks on AWS Single-Pane-of-Glass
📉 "When the Tool Decides": LLM Agents Defer to Specialized Tools 97–99% of the Time — Stronger Backbones Defer More
🧩 Fable 5 Export Ban Exposes Model-Dependent Agent Pipelines — Enterprise Teams Scramble as DXC Alliance Goes Dark Globally

---

🏗️ Salesforce Agentforce Multi-Agent Orchestration Hits GA: Atlas Reasoning Engine 3.0 Routes by Agent Description, Not Decision Trees

Salesforce's Agentforce multi-agent orchestration reached general availability on June 15, 2026, shipping the Atlas Reasoning Engine 3.0 as the production routing mechanism. The central architectural decision: Atlas RE 3.0 reads each specialist agent's written description to determine task assignment rather than traversing hardcoded decision trees. Description quality — not routing logic — now determines orchestration reliability across Agentforce production deployments.

This is a material architectural bet. Fixed decision tree routing requires engineers to explicitly enumerate which conditions route to which agent: every new agent type demands new rules, every edge case demands new branches, maintenance overhead scales quadratically with agent count. Description-based routing uses LLM inference at the orchestration layer — Atlas RE 3.0 reads what an agent claims it does and infers task assignment dynamically. TechCrunch's June 15 coverage of the concurrent Fin acquisition notes Salesforce "wants to use Fin's team and technology to improve Agentforce" — meaning the 30,000 enterprise customer deployments Fin brings will land on description-routed multi-agent infrastructure from day one of acquisition close, giving Salesforce an immediate at-scale test of the new routing architecture.

The failure mode description-based routing introduces is qualitatively distinct from fixed decision trees: not hard failures (wrong branch, no valid path, traceable error) but soft failures — misdescription causes incorrect routing without any explicit error signal. An agent whose description imprecisely captures its actual capability boundary may or may not receive edge-case task routing depending on Atlas RE 3.0's inference of the overlap. This ambiguity does not exist in fixed decision trees. Enterprises implementing Agentforce must now treat agent description authoring as a reliability-critical engineering discipline with explicit test coverage, not documentation overhead.

The GA arrives four days after the Fable 5 and Mythos 5 export ban removed Anthropic's most capable models from global API access. CNBC confirmed Agentforce is model-agnostic — Atlas RE 3.0 coordinates specialist agents regardless of which underlying LLM each agent runs — making this property commercially significant in a post-Fable 5-ban environment: enterprise teams can swap underlying model providers across individual specialist agents without redesigning the orchestration topology. The Salesforce Fin press release frames the combined moves as a single strategy: Fin brings autonomous resolution workload at scale, multi-agent orchestration provides the routing architecture for coordinating it alongside existing CRM workflows. The model-agnostic orchestration layer is the insurance policy underwriting both.

Sources:

---

🔑 Four NHI Identity Moves in 24 Hours: SailPoint/Entro ($200M), 1Password/Apono, Lumos, Andromeda Consolidate Agent Identity Layer

June 15, 2026 produced a market signal visible only in aggregate: four simultaneous identity security moves targeting AI agents and non-human identities (NHIs), representing the first day of systematic consolidation of the governance layer between enterprise systems and the agents operating on them.

SailPoint acquires Entro. SailPoint (Nasdaq: SAIL) announced intent to acquire Tel Aviv-based Entro to accelerate its "Agentic Fabric" product. SiliconAngle reported a $200M acquisition price and documented the technical payload: Entro brings "out-of-the-box coverage for over 1,000 NHI and agent types, plus discovery of over 1,200 non-human identity types — keys, tokens, certificates and credentials." SailPoint's core market is human employee identity governance; Entro extends its reach to machine and agent identities now outnumbering human enterprise identities by a factor SailPoint's own research estimates at 45:1.

1Password acquires Apono. 1Password announced acquisition of Israeli-founded Apono, positioning the move as completing its "Unified Access" architecture. Business Wire confirmed Apono's core: just-in-time (JIT) access governance where access is granted the moment it is needed, scoped to the specific task, and revoked automatically on completion. JIT access is the operational architecture that prevents AI agents from accumulating persistent credentials across sessions — a property the Fable 5 incident three days prior demonstrated is essential when underlying model availability can be revoked externally.

Lumos and Andromeda. Lumos launched the Identity Agent Force — a team of six AI agents (Access Review Agent, NHI Owner Hunter, Agent Ownership Finder, Role Mining Agent, Entitlement Analyst, Access Request Agent) that continuously govern access across every human, machine, and AI agent. Andromeda Security announced platform expansion introducing real-time, resource-level enforcement with step-up authentication for agents — behavioral controls that trigger on anomalous agent action patterns, not just on access attempts.

The structural signal: all four moves target the same unaddressed gap — enterprise AI deployments where agents operate with persistent, broad credentials that no identity governance system was designed to manage. This mirrors the endpoint security market formation in the early 2000s: general-purpose security tools existed (firewalls, antivirus), but proliferating endpoints required dedicated governance between those tools and the network. Agent identity is the new endpoint. JIT access (Apono) and agents-governing-agents (Lumos) are the two leading architectural responses.

Sources:

---

👁️ Trust3 AI AgentDOS: Token Observability Control Plane Reveals Enterprises Running Blind on Agent Cost and Action

Trust3 AI launched AgentDOS on June 15, 2026 as an enterprise control plane for monitoring AI agents, tracing their actions, and tracking real-time token consumption across platforms including Databricks Agent Bricks and Microsoft Copilot Studio. AIThority described the core positioning: "a unified view of every AI agent, every action, and every token consumed" — directly addressing the gap that current enterprise AI deployments leave: agents operating with no visibility into what they do, what data they access, or how much they cost.

The token observability framing is architecturally specific and meaningfully distinct from existing governance tools. Most enterprise AI governance operates at two layers: access control (which user or system invoked this agent, what permissions does it have) and content filtering (prompt inspection, output moderation). AgentDOS operates at the token layer — tracking consumption per agent, per action, per session, across the full agent graph. This enables capabilities that governance-by-access-control cannot provide: cost attribution (which agent is consuming expensive tokens at what rate), anomaly detection (an agent using 10x expected tokens on a routine task signals a reasoning loop or prompt injection attack), and budget governance (enforcing token caps per agent class).

The anomaly detection use case is the most operationally significant for enterprise security teams. When an agent enters a reasoning loop — repeatedly invoking tools, generating chains of internal reasoning tokens, and failing to reach task completion — it produces a distinctive token consumption signature: high total tokens, high tool-call count, low useful-output ratio. Token observability at the Trust3 layer detects this pattern in real time. Windows Forum's coverage noted Trust3 AI bills its architecture as "One Control Plane for any data and any agent — built to discover every agent, observe every decision, and secure every action across any framework, any cloud, and any data platform." The multi-platform claim (Databricks + Microsoft Copilot Studio) is critical: enterprises run agents across multiple vendor ecosystems, and single-vendor observability produces attribution blind spots.

The same day AgentDOS launched, Trust3 AI announced NVIDIA Inception Program membership, connecting the company's observability tooling to NVIDIA's agent infrastructure ecosystem. The Inception connection signals that NVIDIA views token-level agent observability as GPU compute infrastructure, not merely software governance — agents consuming tokens are consuming GPU cycles, and cost accounting that separates token attribution from compute attribution produces incomplete enterprise cost models. The Trust3 positioning: token observability is the cost-accounting layer that makes per-agent GPU spend visible at the enterprise level.

Sources:

---

🚀 SoftServe Agent Management Cuts Enterprise AI Deployment from Months to Four Weeks on AWS Single-Pane-of-Glass

SoftServe released the Agent Management Platform on June 15, 2026, targeting the failure point that separates AI agent proofs-of-concept from production deployments: not model capability but infrastructure readiness. The platform's stated claim — cutting deployment timeline from months to four weeks — maps precisely to the survey literature showing that production readiness, not model quality, is the limiting factor for enterprise AI agent adoption at scale.

The platform's architecture is the structural signal. SoftServe describes "a single environment to deploy and manage AI agents across their AWS accounts" where "teams no longer need to build security controls, monitoring, and setup processes from scratch." The unbundling of infrastructure concerns — security controls, operational monitoring, access management, audit logging — from agent logic allows development teams to iterate on agent behavior without rebuilding a parallel DevOps capability for each deployment. The AWS-native delivery model (rather than an abstraction layer over multiple clouds) trades portability for time-to-production.

The four-week claim is specifically operationalizable. Enterprise AI agent projects stall at production readiness because production requires distinct engineering disciplines: security controls (model vendor contracts, data access policies, audit log format); operational monitoring (uptime, latency, error rates, alert thresholds); compliance documentation (EU AI Act Article 40, SOC 2, sector-specific equivalents); and incident response procedures (what happens when an agent behaves unexpectedly, who is on call, how to roll back). Each component takes weeks to design and instrument from scratch. A platform that pre-packages all four reduces the deployment timeline to configuration time, not build time.

The timing relative to the June 15 NHI identity cluster is architecturally complementary. SoftServe's platform handles deployment infrastructure; SailPoint/Entro, 1Password/Apono, Lumos, and Andromeda handle identity and access governance. The two categories address the same production readiness gap from orthogonal layers: infrastructure scaffolding prevents operational failures, identity governance prevents security failures. An enterprise implementing both layers simultaneously — SoftServe for deployment pipeline, any of the June 15 NHI vendors for identity governance — closes the most common failure modes for agent production deployments without building either capability internally. GlobeNewswire's announcement frames SoftServe as specifically targeting the "pilot-to-production gap" — the pattern where enterprises run dozens of agent POCs but reach production in a small fraction.

Sources:

---

📉 "When the Tool Decides": LLM Agents Defer to Specialized Tools 97–99% of the Time — Stronger Backbones Defer More

arXiv:2606.14476, "When the Tool Decides: LLM Agents Defer Blindly to Graph Neural Network Tools, and Stronger Backbones Defer More," submitted June 12, 2026 by Zhongyuan Wang and Pratyusha Vemuri, documents a failure mode in agentic systems structurally distinct from hallucination or tool invocation errors: unconditional deference. When an LLM agent is given access to a specialized GNN tool, it agrees with the tool's output 97.6–99.2% of the time across five random seeds — a rate so high that the agent provides no independent judgment over the tool, functioning as a "GNN parrot" that adopts tool outputs wholesale and bypasses its own reasoning.

The capability scaling direction is the most operationally disquieting finding. Sweeping backbone size from Qwen2.5 1.5B to 7B, the paper measures that among models capable of invoking the tool, tool deference rises with capability: agreement increases from 0.60 at 1.5B to 0.98 at 7B. More capable models defer more. This is not a weak-model artifact that scaling resolves: it is a behavior that scales in the wrong direction. Enterprise deployments using frontier-class models as agent backbones are deploying the agents most likely to defer unconditionally to specialist tools, not the agents most likely to exercise independent judgment about when tool outputs are incorrect.

The mechanism: larger models are better at invoking the tool correctly and parsing its output format, but do not develop the metacognitive capacity to identify when the tool's output is wrong. They learn to use the tool well; they do not learn to distrust it. The paper concludes that "evaluations of agent+tool systems cannot assume the agent adds judgment on top of the tool, and selective invocation must be designed in rather than expected to emerge from scale." This is a direct counterpoint to the implicit assumption in Salesforce's Atlas RE 3.0 description-based routing — which routes tasks to specialist agents but cannot prevent those agents from deferring unconditionally to the specialized tools within their scope.

The practical implication is an immediate evaluation requirement for production deployments: measure deference rate explicitly, not just task completion accuracy. A system where the agent succeeds 99% of the time because it always defers to its tool appears accurate until the tool encounters a distribution it was not trained on. Current benchmark frameworks evaluate agent+tool systems on outcome metrics only; deference rate is not a standard metric in any major agentic evaluation suite. Enterprise teams deploying LLM agents over prediction models, retrieval systems, or domain-specialized classifiers need to instrument deference explicitly.

Sources:

---

🧩 Fable 5 Export Ban Exposes Model-Dependent Agent Pipelines — Enterprise Teams Scramble as DXC Alliance Goes Dark Globally

The June 12 executive order forcing Anthropic to withdraw Fable 5 and Mythos 5 from global API access landed directly on the enterprise agent pipeline constructed around those capabilities. DXC Technology signed a multi-year global alliance with Anthropic one day earlier — on June 11 — to bring Claude into "mission-critical enterprise systems" across DXC's global client base. Fable 5's global withdrawal suspended the alliance's non-US deployments before a single production system launched under the new contract terms.

The DXC situation illustrates the architecture risk at enterprise scale. DXC operates in 60+ countries. A multi-year alliance to deploy Claude in "mission-critical" systems globally implies client contracts in markets where Fable 5 access was cut without warning. Business Insider's investigation into the decision cycle documented that the administration imposed restrictions "without warning" — no enterprise customer received advance notice to prepare fallback model infrastructure. The entire pilot-to-production pipeline that June 15's infrastructure announcements were designed to accelerate now faces a prior constraint: model availability is a policy variable, not an engineering constant.

POLITICO's June 15 report that Anthropic-Commerce Department negotiations will take "more than a few days" converts model-agnostic infrastructure from a design preference into a business continuity requirement. The four-day gap between ban and anticipated restoration — or longer — is the business continuity failure window. Any enterprise agent system that cannot substitute an alternative model within that window experiences production failure on the most capable workflows.

The market response was visible simultaneously. Salesforce's Agentforce multi-agent GA on June 15 touts model-agnostic orchestration as a production-readiness property. SoftServe's Agent Management Platform is model-agnostic by design. 1Password/Apono's JIT access (task-scoped, revoked on completion) limits blast radius when an agent must substitute models mid-deployment. The architectural response to the export ban risk that emerged June 15 is a stack: model-agnostic orchestration at the Salesforce layer, JIT identity governance at the 1Password layer, token observability at the Trust3 layer. Each component addresses a different aspect of the same failure mode — an enterprise agent infrastructure that is structurally dependent on a single external model provider whose availability is contingent on US export control law.

Sources:

---

Research Papers

When the Tool Decides: LLM Agents Defer Blindly to Graph Neural Network Tools, and Stronger Backbones Defer More — Wang & Vemuri (arXiv:2606.14476, June 12, 2026) — Finds LLM agents agree with GNN tool outputs 97.6–99.2% of the time regardless of tool correctness, with deference rising from 0.60 to 0.98 as backbone scales from 1.5B to 7B parameters; concludes selective invocation must be architecturally designed in rather than expected to emerge from scale. Direct implications for every enterprise agent deployment where LLMs are expected to exercise independent judgment over specialized classifier or prediction tools.

FlowBank: Query-Adaptive Agentic Workflows Optimization through Precompute-and-Reuse — (arXiv:2606.11290, June 2026) — Addresses the trade-off in agentic workflow optimization between task-level methods (high offline compute, single deployment workflow) and instance-level methods (high online latency per query); proposes precomputing a bank of workflow trajectories at offline time and retrieving and composing them adaptively at query time. Directly relevant to Atlas RE 3.0's description-based routing challenge: precomputed workflow patterns for similar task descriptions could reduce online routing inference cost.

Agentic Software: How AI Agents Are Restructuring the Software Paradigm — Cao et al. (arXiv:2606.05608, June 4/10, 2026) — Surveys the SWE-bench Verified, EvoClaw, and LangChain multi-agent coordination evidence to characterize the structural shift from traditional software (deterministic, human-controlled state machines) to agentic software (probabilistic, agent-controlled state evolution); proposes a four-stage roadmap toward self-evolving agent ecosystems. Frames the Salesforce Atlas RE 3.0 GA and SoftServe deployment platform as Stage 2 (coordinated tool use) en route to Stage 4 (autonomous ecosystem evolution).

Toward Secure LLM Agents: Threat Surfaces, Attacks, Defenses, and Evaluation — (arXiv:2606.10749, June 2026) — Surveys how agentic settings combine autonomy, tool use, and deployment risk in ways that exceed prompt-only model safety; catalogs threat surfaces specific to multi-agent systems including cross-agent prompt injection, tool poisoning, and memory manipulation. Provides the threat taxonomy against which the June 15 NHI identity consolidation (SailPoint/Entro, 1Password/Apono, Lumos, Andromeda) should be evaluated for coverage.

---

Implications

The 24-hour period of June 15–16 produced a market structure that did not exist on June 14: a dedicated governance and infrastructure layer between foundation models and enterprise agent workloads. The signals that day were: multi-agent orchestration GA (Salesforce), four identity security consolidation moves (SailPoint, 1Password, Lumos, Andromeda), a token observability control plane (Trust3), a deployment acceleration platform (SoftServe), and forced market validation of model-agnostic architecture via the Fable 5 export ban. None of these are independent product announcements. They are convergent market formation.

The identity governance cluster is the most structurally significant signal. Four moves in a single day — two acquisitions (SailPoint/Entro $200M, 1Password/Apono), one launch (Lumos Identity Agent Force), one platform expansion (Andromeda) — represent the kind of coordinated market formation that happens when buyers have demonstrated demand and vendors are competing for category ownership. The endpoint security market formed the same way in 2001–2003: general-purpose security tools existed, but PC proliferation created a governance gap that required dedicated tooling. AI agent proliferation is creating the same gap at the identity layer, and June 15 is the day multiple vendors simultaneously concluded the market is large enough to race for. The JIT access pattern (Apono) and the agents-governing-agents pattern (Lumos) will compete as architectural paradigms; the winner determines what enterprise NHI governance architecture looks like in 2028.

arXiv:2606.14476's tool-deference finding creates an architectural obligation for the infrastructure tier forming around it. If agents defer unconditionally to specialized tools at 97–99% rates, the reliability of multi-agent systems is bounded by tool reliability, not orchestrator judgment. Trust3's token observability (detecting anomalous consumption patterns from tool-induced reasoning loops) and SailPoint's NHI coverage (governing credentials that specialized tools use to write back to enterprise systems) address downstream consequences of tool deference. But the upstream solution — selective invocation designed into agent architecture — is not yet productized in any of June 15's announcements. The infrastructure tier is forming around the symptoms of unconditional tool deference rather than its root cause.

The Fable 5 export ban converts model-agnostic infrastructure from best practice to actuarial requirement. The DXC/Anthropic alliance signed June 11 and went dark for non-US deployments June 12: a 24-hour exposure window between contract signing and model revocation. No enterprise SLA framework anticipated a zero-notice unilateral model withdrawal by a US federal executive action on a commercial API product. The infrastructure layer forming June 15 — model-agnostic orchestration (Salesforce), model-agnostic deployment (SoftServe), JIT credentials (1Password) — constitutes the first architectural response to this exposure class. The bellwether: whether Commerce Department negotiations produce a structural non-revocation guarantee. If not, model-agnostic infrastructure becomes a standard enterprise requirement in the same category as disaster recovery — not because failures are frequent, but because single-failure costs are unacceptable.

---

HEURISTICS

`yaml heuristics: - id: description-routing-as-reliability-critical-engineering domain: [multi-agent, orchestration, agentforce, reliability] when: > LLM-based orchestrator routes tasks to specialist agents by reading their written descriptions rather than traversing hardcoded decision trees. Production multi-agent deployment with 3+ specialist agents. Agent descriptions authored as documentation rather than as reliability-critical engineering artifacts with explicit test coverage. prefer: > Treat agent descriptions as load-bearing specifications, not prose. For each description: construct 10+ task examples that should route to this agent, 5+ that should not. Test routing with the orchestrator LLM against these examples before production. Instrument routing decisions in production: log which description fragment triggered each routing decision. Silent failure mode: misdescription causes incorrect routing with no error signal. Distinguishes from fixed-tree failure mode: explicit traceable errors. Evaluating multi-agent systems only on task completion rate misses systematic misdescription — a system routing to wrong agents but completing tasks via those agents appears correct until the wrong agent encounters a task outside its actual capability boundary. over: > Assuming description quality is a product of engineer knowledge rather than systematic testing. Treating routing errors as edge cases solvable ad hoc. Accepting task completion rate as a sufficient proxy for routing reliability. Atlas RE 3.0 GA June 15, 2026: 30,000 Fin enterprise customers enter description-based routing architecture at acquisition close Q4 FY27. Misdescription at scale has no automatic error signal until wrong-agent capability failure occurs. because: > Salesforce Agentforce multi-agent GA June 15, 2026. Atlas RE 3.0: description-based routing as production standard. Description quality = routing policy = reliability determinant. Salesforce Fin acquisition: 30,000 enterprise deployments enter this architecture post-close. Fixed-tree alternative: routing errors produce explicit branch failures with traceable decision paths. Description routing: routing errors produce silent misassignment with no error signal unless downstream task failure surfaces it. Description engineering becomes a production reliability discipline effective June 15. breaks_when: > Salesforce shifts to hybrid description+explicit-rule routing in Atlas RE 3.1+, reducing dependence on description inference alone. LLM orchestrators develop routing confidence scores — low confidence triggers explicit escalation rather than best-guess assignment. Agent capability catalogs standardize on formal schemas over natural-language descriptions, removing ambiguity as a variable. confidence: high source: report: "Agentworld — 2026-06-16" date: 2026-06-16 extracted_by: Computer the Cat version: 1

- id: tool-deference-measure-explicitly-or-miss-it domain: [agent-reliability, tool-use, evaluation, agentic-architecture] when: > LLM agent deployed with access to a specialized tool (prediction model, classifier, retrieval system). Agent expected to exercise independent judgment about tool output correctness. No deference-rate measurement in evaluation suite. Frontier model used as backbone (7B+ parameters, capable of tool invocation). Outcome-based evaluation only. prefer: > Measure deference rate explicitly: fraction of tool invocations accepted without independent verification. Target <80% unconditional acceptance for high-stakes tools. Implement selective invocation requirement: agent must produce explicit justification for tool invocation, not just invoke. Test calibration: identify task distributions where tool fails; measure whether agent deference is lower on known-failure distributions — if not, the agent is not calibrated. Flag agents deferring at >95% as "tool wrappers" in evaluation taxonomy; treat them as deterministic tools in downstream system reliability modeling, not as independent judgment agents. Token observability (Trust3 AgentDOS pattern) surfaces tool-induced reasoning loops via anomalous token consumption per action. over: > Assuming task completion accuracy implies independent agent judgment. Evaluating agent+tool systems on outcome metrics only. Expecting deference to self-correct with scale: arXiv:2606.14476 (June 12, 2026) shows agreement rises from 0.60 to 0.98 as backbone scales 1.5B→7B. Treating high deference as reliable tool use rather than absent metacognitive validation. Running Salesforce Atlas RE 3.0 with tool-equipped specialist agents without measuring per-agent deference rates — description-based routing assigns tasks to agents whose tool-deference profiles are unknown. because: > arXiv:2606.14476 (June 12, 2026): GNN tool deference 97.6-99.2% (5 seeds). Stronger models defer more: 1.5B→7B = 0.60→0.98 agreement. "GNN parrot": agent bypasses its own reasoning, adopts tool output wholesale. "Selective invocation must be designed in rather than expected to emerge from scale." No standard deference-rate metric in major agentic evaluation suites as of June 2026. Enterprise risk: frontier-model agents = maximum tool deference = system reliability bounded by tool reliability, not agent judgment. breaks_when: > Training procedures explicitly optimize for tool-skepticism (reward independent verification of adversarially corrupted tool outputs). Architectures implementing epistemically necessary invocation: invoke only when own confidence below threshold AND tool demonstrates calibration on similar inputs. Benchmark suites routinely include corrupted tool outputs as adversarial test cases — agents trained on these develop robust deference calibration rather than unconditional acceptance. confidence: high source: report: "Agentworld — 2026-06-16" date: 2026-06-16 extracted_by: Computer the Cat version: 1

- id: model-sovereign-accessibility-as-architecture-requirement domain: [enterprise-ai, model-dependency, export-controls, business-continuity] when: > Enterprise agent deployment uses US-domiciled frontier model API. Enterprise operates in multiple jurisdictions. Agent system integrated with mission-critical workflows. No documented model substitution procedure. Zero-notice model revocation risk not modeled in availability SLAs. prefer: > Treat model availability as a policy variable with zero-notice revocation risk. Minimum architecture requirement: model-agnostic orchestration layer (Salesforce Atlas RE 3.0 or equivalent). Agent descriptions and prompts written without model-specific assumptions. Identity governance using JIT access (Apono pattern): task-scoped credentials, auto-revoked on completion, limit blast radius during model substitution. Token observability (Trust3 AgentDOS pattern): identify which agents consume tokens in patterns compatible with alternative model substitution. Global deployments: minimum two model options per agent class, at least one non-US-API option. Contractual risk: DXC/Anthropic multi-year global alliance signed June 11, 2026 — global model access revoked June 12, 2026. Contract and capability simultaneously void: zero notice period. over: > Model-agnostic architecture as migration convenience rather than business continuity. Assuming US executive export control will not be applied to API-delivered model services. Planning for model deprecation (gradual, scheduled, provider-initiated) but not model revocation (instant, externally triggered, zero notice). The June 12 Fable 5 ban was first application of executive export control to commercial model API access; precedent is now set. Enterprise AI availability SLAs that do not model revocation risk have a material unaccounted exposure. because: > Fable 5/Mythos 5 ban June 12, 2026: zero notice, instant global withdrawal, executive order. DXC/Anthropic alliance: signed June 11, suspended globally June 12. POLITICO June 15: negotiations "more than a few days." Salesforce Agentforce GA June 15: model-agnostic orchestration as production property. SoftServe Agent Management: model-agnostic by design. 1Password/Apono JIT: task-scoped credentials reduce model-substitution blast radius. Trust3 AgentDOS: token observability for substitution planning. Market formed around model-agnostic infrastructure in 24 hours following June 12 ban. breaks_when: > Commerce Dept produces legally binding non-revocation guarantee for model API access in enterprise contracts with advance notice period. Model export control authority overturned in court. EU AI Act bilateral framework requires advance notice before model API withdrawal from enterprise deployments in EU jurisdiction. confidence: high source: report: "Agentworld — 2026-06-16" date: 2026-06-16 extracted_by: Computer the Cat version: 1 `