The Coordination Gap: Intent Governance Is the Layer Nobody Built

The Insight

The governance conversation in 2026 is split across two layers. At the action layer, Microsoft's Agent Governance Toolkit intercepts every tool call before it reaches the wire — deterministic, sub-millisecond, covering all ten OWASP agentic AI risks.[1] At the memory layer, Andrej Karpathy describes building markdown wikis and CLAUDE.md files so agents don't rediscover context from scratch on every session.[2] Both are real problems. Both are being addressed. Neither addresses what happens between agents — when one pipeline stage hands its findings to the next and the constraint that was declared as critical arrives as a suggestion.

Only 21% of organizations report having a mature model for agent governance, and the gap compounds as agent networks grow.[3] The multi-agent coordination problem is not a model quality problem. Frontier models are capable. The failure mode is structural: when two agents operate from isolated memory stores with different definitions of the same term, the supervisor has no mechanism to resolve the conflict on principled grounds.[4] The intent evaporated at the handoff. Not because the model failed — because the architecture provided no container for the intent to travel in.

The EU AI Act's high-risk AI system requirements take effect August 2, 2026 — mandating traceability, human oversight, technical documentation, and audit trails for every consequential AI decision.[5] The regulation does not distinguish between what an agent does and what an agent means. Proving to a regulator that an agent operated lawfully requires demonstrating that the intent authorizing each action survived the pipeline intact. Runtime policy enforcement catches unauthorized actions. It cannot prove authorized intent traveled correctly.

Five tools built independently for different problems — RECALL (publishing), Wake Intelligence (temporal memory), EMBER (agent artifact language), Rune Protocol (reactive binding), and Mere (workbook format) — converged on the same three-property structure at their coordination boundaries: immutability (constraints cannot be softened in transit), addressability (every constraint is queryable by downstream agents), and provenance (the origin and authorization of every constraint is permanently traceable).[6] The convergence was not by design. Each tool was solving its own problem. The primitive emerged from the intersection. Convergent discovery across five independent implementations is the strongest possible evidence that the structure is architecturally necessary — not a design preference.

The Two Governance Layers Already Built

The action layer and the memory layer have received the industry's attention. Microsoft AGT governs what agents are permitted to do — blocking destructive operations, enforcing identity, sandboxing execution.[1] The toolkit covers all ten OWASP agentic AI risks with deterministic enforcement that makes unauthorized actions structurally impossible, not merely unlikely. Karpathy's wiki pattern governs what agents remember — compiled markdown that survives session resets, reducing the rediscovery cost of context across runs.[2] These layers are necessary. They are not sufficient. The third layer — what intent means as it crosses agent boundaries — remains structurally unaddressed by any major framework. Microsoft AGT's own open issues include a blocked RFC on governing agent skills across context injection boundaries, and an unresolved feature request for human-in-the-loop approval chains in the policy evaluator. The action layer cannot govern what it cannot see — and intent drift happens before any action is taken.

The key risk is not intelligence. It is coordination failure at scale. — Multi-Agent Orchestration in 2026, Medium

The Primitive That Emerged From Practice

A governance primitive is not a framework applied to a domain. It is what the domain is built from. The distinction matters because frameworks are adopted — primitives are discovered. The three-property structure (immutability, addressability, provenance) appeared independently across five production systems solving different problems in different domains. RECALL embeds source and authorship in every compiled artifact — the AUDIT DIVISION enforces provenance as a compile-time constraint, not a logging afterthought.[6] Wake Intelligence stores causal memory with WHY chains that survive agent session resets. EMBER carries artifact state across a seven-agent legacy modernization pipeline in .sil files that every downstream agent reads without prior context. Rune Protocol's ? sigil makes intent machine-readable, co-located with the code it governs, version-controlled, queryable at runtime. Mere workbook format carries screens, state, and behavior as a single sovereign artifact. The pattern in all five: the constraint travels with the artifact. The intent is not reconstructed at each handoff — it is read from a structure that carries its own authorization. That is the primitive. The EU AI Act's August 2026 enforcement deadline makes it urgent. The 79% of organizations without mature agent governance make it commercially relevant. The five-tool convergence makes it structurally necessary.[5][7]

The authorship dimension is the governance primitive's most concrete consequence. In a multi-agent pipeline, a human-approved risk threshold and an agent-inferred optimisation threshold may appear as identical structured data in the same context package. Without structural authorship, they are indistinguishable. Three tools independently addressed this: RECALL enforces CREATED-BY — Human, AI compositor, or AI agent — at compile time; a document without an authorship declaration does not compile.[6] Rune Protocol carries metadata.author on every ? binding annotation in rune.schema.json. Wake Intelligence stores authorType on every saved context, surfaced via the auditor personality mode — which groups retrieved contexts by author class, immediately distinguishing load-bearing human constraints from negotiable agent optimisations.[6] UC-234 extends the Semantic Intent as Single Source of Truth framework established in 2025[11] — applying the same WHAT+WHY governance contract from the two-party developer-AI context to the multi-agent pipeline.

-- UC-234: The Coordination Gap
-- Semantic Intent as Governance Primitive for Agentic Systems
-- Sense → Analyze → Measure → Decide → Act

FORAGE agent_pipelines
WHERE intent_drift_at_handoff = true
  AND action_governance_present = true
  AND intent_governance_absent = true
  AND eu_ai_act_enforcement < '2026-08-02'
ACROSS D5, D4, D6, D2, D1, D3
DEPTH 3
SURFACE coordination_gap

DIVE INTO governance_layers
WHEN action_layer_coverage = 'complete'
  AND memory_layer_coverage = 'partial'
  AND intent_layer_coverage = 'none'
TRACE primitive_emergence
EMIT coordination_gap_signal

DRIFT coordination_gap
METHODOLOGY 85    -- governance frameworks exist (OWASP, EU AI Act, NIST)
PERFORMANCE 35    -- intent layer structurally unaddressed by all major toolkits

FETCH coordination_gap
THRESHOLD 1000
ON EXECUTE CHIRP critical 'Intent governance primitive discovered through convergent emergence across 5 independent production systems. EU AI Act enforcement deadline August 2026. 79% of organizations remain ungoverned at the intent layer.'

SURFACE analysis AS json

Key Insights

Related Cases