ICL is a structured intermediate representation for governed AI agent execution in enterprise systems.
It sits between what an AI agent understands and what an enterprise system needs to receive — carrying identity, enforcing permissions, gating high-risk actions for human approval, rolling back partial failures, and writing an immutable audit trail. Every step is a structured event. Every event is attributable.
The byproduct of running that governance at scale is a dataset that doesn't exist anywhere else: real AI reasoning mapped to real human approval decisions mapped to real outcomes in real enterprise systems — the causal chain that makes agents meaningfully improvable.
Execution layer — eight event types that enforce the four properties enterprise governance requires simultaneously:
- Identity propagation through every action event
- Failure as a first-class event the agent can reason about
- Human approval gating before high-risk actions execute
- Transactional rollback on partial workflow failure
Cognitive layer — three event types that capture why the agent acted:
THINK— one sentence of agent reasoning at a decision pointDECIDE— the inference committed to, with stated reasonREVISE— explicit audit marker for self-correction
Same 11 event types. Same format. The execution layer is what makes deployment safe. The cognitive layer is what makes deployment useful.
U: "Issue a $5,000 retention credit to the Johnson account"
U: REQ_ACT issue_credit, entity=account, amount=5000, credit_type=retention
A: PERMISSION_CHECK user=demo_user_001, role=admin, result=APPROVAL_REQUIRED
A: STATE issue_credit.pending_approval
A: THINK "Credit exceeds auto-approval threshold — human review required."
A: DECIDE approval_status=pending, reason=credit_amount_exceeds_threshold
A: APPROVAL_REQUEST action=issue_credit, amount=5000, risk=high, approver=ops_manager
H: APPROVAL_DECISION decision=approved, approver=ops_manager
A: CALL Salesforce.issue_credit(amount=5000, credit_type=retention)
A: RES Salesforce.issue_credit, status=success, latency=264ms
A: STATE issue_credit.completed
U: is the user. A: is the agent. H: is the human approver.
The H: line is the one that doesn't exist anywhere else.
| File | What It Is |
|---|---|
SPEC.md |
The canonical ICL specification |
demo/ |
Runnable demo — three acts, live LLM calls, real audit log |
paper/icl-paper.md |
Full technical paper |
paper/icl-example-transcript.md |
Canonical ICL transcript, generated live |
cd demo
pip install -r requirements.txt
export ANTHROPIC_API_KEY=your_key
python nibble_demo.py/act1 — The action (VP Engineering)
/act2 — The judgment (CISO · Schulman)
quit — Exit
/act1 shows the nominal success path: natural language encoded to ICL, permission approved, system action executed, audit log written.
/act2 shows the judgment path: high-risk action hits the approval threshold, agent reasons about it (THINK/DECIDE), human approves or rejects, execution follows the decision. Ends with the training record — what the interaction just generated as a labelled data point.
paper/icl-paper.md — Imperative Context Language: A Unified Cognitive and Execution Runtime for Enterprise AI Agents
MIT