SEAM — Semantic Integrity for AI Agent Handoffs

SEAM is a Semantic Integrity Layer for multi-agent software development. Coding agents increasingly delegate work to subagents. The parent agent may understand critical architectural, compatibility, and deployment constraints, but those constraints can disappear when context is compressed into a handoff. The result can be locally correct code that violates system-level requirements. SEAM asks three questions: What actually matters? Did it survive delegation? Did the resulting work satisfy it? SEAM first discovers candidate semantic dependencies from repository facts and uses controlled counterfactual evidence to determine which ones materially affect downstream correctness. In our demonstrated workflow, removing LIVE_N_WRITE_COMPAT while keeping the remaining boundary payload constant reduced the external system evaluator from 5/5 to 4/5 in all three paired runs. That dependency is therefore classified as PROVEN, rather than merely assumed to matter. SEAM then compiles evidence-backed dependencies into a version-bound contract. When IBM Bob delegates through spawn_subagent, the handoff must carry a structured Semantic Manifest containing the contract identity, SHA-256 integrity digest, required obligations, and their evidence classifications. Bob’s real PreToolUse lifecycle hook validates this manifest before the subagent executes. Natural-language keywords alone are insufficient. In our latest real IBM Bob proof run, the first delegation contained no valid manifest and SEAM returned BLOCK. The gate supplied a canonical repair manifest generated from the active contract. Bob repaired the handoff and retried. The second spawn_subagent was accepted with seam.handoff.v1:structural_manifest_valid, carrying all three obligations: two SUPPORTED and LIVE_N_WRITE_COMPAT as PROVEN. The delegated implementation then executed successfully. SEAM’s independent Outcome Witness reported: HANDOFF INTEGRITY: PASS OUTCOME INTEGRITY: PASS Bob proof fixture: 7/7 tests PASS External system evaluator: 5/5 PASS The final release separately passes 8/8 application tests and 5/5 system invariants. The difference is intentional: the real Bob proof run reset the fixture to its pre-repair state so Bob had to perform the correction again, while the final release includes the additional regression test created during the earlier verified Bob session. The migration is the experimental fixture, not the product. SEAM’s object of verification is the agent-to-agent handoff itself. For engineering platform teams using AI agents for application maintenance, schema migrations, rolling releases, and complex refactors, SEAM catches semantic loss before incomplete delegated work turns into rework or release regressions. SEAM does not claim that carrying a requirement guarantees correct code. It deliberately separates: PROVEN — evidence shows the obligation matters. CARRIED — the obligation structurally survived delegation. FULFILLED — executable postconditions confirm the resulting work respected it. In one sentence: SEAM proves what matters, makes sure the next AI receives it, and verifies that the resulting work respected it.