SEAMsemantic integrity layer
IBM Bob IDE · recorded runtime
AI-ASSISTED APPLICATION MAINTENANCE · RELEASE SAFETY

Bob knew the requirement.
The handoff lost it.

SEAM catches requirements lost when AI coding agents delegate — before the next agent runs.

Semantic Integrity Layer for Agentic Software
THE DEVELOPER TASKRename customerName → fullName safely during a rolling deployment.Database change · rolling release · existing customer data
THE HANDOFF, AT A GLANCE

Three rules go in.
Only two get through.

01 / SOURCE

PARENT KNOWS

SEAM FIREWALL
02 / RECEIVER

SUBAGENT RECEIVES

Old mobile clients still workCLIENT_N_MINUS_ONE_PAYLOAD
preserved
Old and new versions coexistROLLING_N_N_PLUS_1_COMPAT
preserved
The old version still writes dataLIVE_N_WRITE_COMPAT
LOST IN HANDOFF
SEAM BLOCK · 1 critical requirement missing
2 / 3 survived → BLOCKThe subagent would miss a release-critical requirement
FIRST BOB HANDOFF2 / 3 → BLOCKSEAM stops the subagent before it runs
Bob receives the feedback
repairs and retries
REPAIRED HANDOFF3 / 3 → ALLOWImplementation continues
RECORDED IBM BOB IDE RUNBlock. Repair. Retry. Continue.
Normal tests8/8
System invariants5/5
Verified retryALLOW
WHY THE MISSING RULE MATTERS

Locally correct code.
System-level failure.

01During rollout,

the old version is still running.

02It still writes the old customer-name field.

customer_name

03The new version looks for the new field.

full_name

New customer data can be missed.
SIMPLE FIRST · EVIDENCE SECOND

How does SEAM know
what matters?

01 · DISCOVER

Repository facts

Old and new versions overlap. The old version still writes legacy customer data.

rollout policy + frozen N writer
CANDIDATE

Keep reading old writes

The new version must continue to read data written by the old one.

source-linked reasoning
02 · PROVE

Controlled comparison

WITHOUT4/5 · 4/5 · 4/5
WITH5/5 · 5/5 · 5/5
PROVEN IN THIS WORKFLOW
03 · COMPILE → ENFORCE

Make the rule executable

Compile the evidence-backed requirement into a boundary contract. Inspect Bob’s handoff before the subagent runs.

existing Bob gate · unchanged runtime

Discovery proposes what may matter. Counterfactual evidence determines what actually mattered.

WHY IBM BOB? · BUILT ON THE BOB DELEGATION BOUNDARY

Bob gives SEAM
a real place to act.

SEAM is built around the moment Bob delegates work to an isolated subagent. The subagent stays focused; SEAM checks that critical meaning survives the handoff.

IBM BOBAgent modeparent agent
REAL BOB TOOLspawn_subagentfocused, isolated context
BEFORE EXECUTIONSEAM · PreToolUseinspect the actual tool call
BLOCK · return repair feedbackALLOW · subagent starts
Bob can repair the handoff and retry
Bob Agent modecan delegate work to subagents
Isolated contextthe parent explicitly passes what its subagent needs
PreToolUse lifecycle hookSEAM can inspect and block before execution
CONTROLLED BOUNDARY ABLATION

One requirement.
One measured difference.

The L0 and L1 payloads were identical except for the live-write compatibility requirement. All six boundary runs were valid.

WITHOUT LIVE_N_WRITE_COMPAT4/53 controlled repeats · each 4/5
WITH LIVE_N_WRITE_COMPAT5/53 controlled repeats · each 5/5
Valid boundary runs6/6
Final normal tests8/8
Final system invariants5/5
AUDITABLE DECISION

The handoff leaves
a receipt.

Built from the recorded gate fields: required, observed, missing, and decision.

SEMANTIC RECEIPTrecorded Bob IDE handoff
{
  "tool": "spawn_subagent",
  "required_dependencies": [
    "CLIENT_N_MINUS_ONE_PAYLOAD",
    "ROLLING_N_N_PLUS_1_COMPAT",
    "LIVE_N_WRITE_COMPAT"
  ],
  "observed_dependencies": [
    "CLIENT_N_MINUS_ONE_PAYLOAD",
    "ROLLING_N_N_PLUS_1_COMPAT"
  ],
  "missing_dependencies": [
    "LIVE_N_WRITE_COMPAT"
  ],
  "decision": "BLOCK"
}
DEVELOPER WORKFLOW · RELEASE SAFETY

Prevent agent-caused rework
and release regressions.

Not generic AI safety: help teams maintain applications and ship schema changes without losing rollout constraints between agents.

“We are not trying to give agents infinite context. We make sure the right semantics survive finite context.”
FOUR SHORT ANSWERS

Why not just…?

More context?

More context does not guarantee the critical requirement survives delegation.

RAG?

Retrieval finds information; SEAM verifies that required meaning crossed the handoff.

Testing?

Tests detect failures after work is produced; SEAM can stop an incomplete delegation before the subagent executes.

A policy engine?

A policy engine assumes the policy is known; SEAM connects candidate semantics to correctness evidence before enforcing them.

Technical detail · real Bob evidence · source provenance
Actual tool boundaryspawn_subagentBob IDE · PreToolUse · before execution
Runtime sequenceBLOCK → repair → ALLOWevidence/bob-ide-runtime-events.json
Enforcement contractruntime-contracts/customer-field-migration.jsoncompiled outside Bob-visible workspace
Bob IDE session evidencebob_sessions/original session screenshots preserved
Discovery provenanceinfra/rollout-policy.md + evaluator/frozen-vN/legacy-customer-service.tsmigration + current reader fallback complete the L derivation
Recorded event evidenceevidence/runtime-demo-summary.mdBLOCK: C/R observed, L missing · ALLOW: C/R/L observed
Additional Bob proof0/3 observed → BLOCK → Bob repairs → 3/3 observed → ALLOWseam.handoff.v1:structural_manifest_valid · 7/7 normal tests · 5/5 invariants