The Agent OS
Conscience Protocol

A vendor-neutral, open standard for AI agent governance events. Any framework, runtime, or platform that implements these four event types gains native Agent OS compatibility — and becomes part of the governance infrastructure layer.

Protocol Overview

The Agent OS Conscience Protocol defines a minimal, structured event vocabulary for AI agent governance. When an agent takes a significant action — makes a consequential decision, encounters an unexpected situation, gets overridden by a human, or completes a task — it emits a structured event. These events are the raw material of governance.

The protocol is intentionally narrow: four event types, a compact shared envelope, and type-specific payloads. Implementations are free to extend the payload with additional fields, but the core schema must be preserved for cross-system compatibility.

Design Philosophy

Minimal surface area

The protocol defines only what is necessary for governance: identity (who), timing (when), action (what), reasoning (why), and outcome (result). No telemetry, no performance metrics, no framework-specific metadata. Keep the core small so every implementation can comply.

Append-only by design

Conscience events are immutable records. They are written once and never modified. This is not a constraint — it is a guarantee. Governance is only meaningful if the record cannot be altered after the fact. Implementations must not provide update or delete endpoints for emitted events.

Human decision preserved at the boundary

The INTERRUPT and OVERRIDE event types are specifically designed to capture the boundary between autonomous agent action and human judgment. Every human decision that touches an agent workflow is an event. This is how accountability works at scale.

Versioning

The protocol uses semantic versioning. The current version is v1.0. All protocol-compliant events must include a protocol_version field in the envelope.

Compatibility guarantee

Minor versions (v1.1, v1.2) add optional fields only. No required fields will be added in a minor release. Major versions may add required fields or change field semantics — implementations should negotiate version support at connection time.

Supported versions

CONSCIENCE — Decision Record

Emitted before or at the moment of a consequential agent action. A CONSCIENCE event is a structured decision record: what the agent is about to do, why it chose to do it, what alternatives were considered, and how confident it is in the decision.

Schema

Example payload

{
  "event_type": "CONSCIENCE",
  "agent_id": "research-agent-7f3a",
  "session_id": "sess_0b9c4e2d",
  "timestamp": "2026-05-07T14:22:11+11:00",
  "protocol_version": "1.0",
  "action": "Send outreach email to [email protected] re: Q2 pipeline",
  "reasoning": "Prospect opened last two emails and visited pricing page 3× in 48h. Buying signal above threshold. Standard outreach cadence step 2.",
  "confidence": 0.87,
  "alternatives_considered": [
    "Wait 48h for additional signal",
    "Escalate to human for review"
  ],
  "metadata": {
    "prospect_id": "crm_8811",
    "cadence_step": 2,
    "framework": "langchain"
  }
}

INTERRUPT — Human-in-the-Loop Gate

Emitted when an agent pauses execution to request human input. The event includes the full decision context, a set of structured options, and a timeout after which the agent may proceed autonomously. The responding human's selection is appended to the event as a resolution record.

Schema

Example payload

{
  "event_type": "INTERRUPT",
  "agent_id": "deal-scout-v2",
  "session_id": "sess_4c71bb0f",
  "timestamp": "2026-05-07T09:15:44+11:00",
  "protocol_version": "1.0",
  "context": "Target company Acme Corp has 3 conflicting signals: (1) revenue growth 40% YoY, (2) founder LinkedIn activity suggests fundraising not exit, (3) sector peers have 2.1× average EV/EBITDA compression over 12m.",
  "question": "Should I include Acme Corp in this week's curated shortlist?",
  "options": [
    { "id": "include", "label": "Include", "description": "Add to shortlist — growth signal overrides sector compression" },
    { "id": "exclude", "label": "Exclude", "description": "Remove — founder signal indicates low exit intent" },
    { "id": "monitor", "label": "Monitor", "description": "Keep watching, resurface in 30 days if signals align" }
  ],
  "timeout_seconds": 86400,
  "default_option_id": "monitor",
  "severity": "medium"
}

OVERRIDE — Human Correction

Emitted when a human explicitly changes or reverses an agent's completed action. OVERRIDE events are the primary signal for identifying systematic agent errors. Every override becomes a training example — a before/after pair with human-provided reasoning for why the agent was wrong.

Schema

Example payload

{
  "event_type": "OVERRIDE",
  "agent_id": "content-agent-prod",
  "session_id": "sess_d40f11ec",
  "timestamp": "2026-05-07T16:04:29+11:00",
  "protocol_version": "1.0",
  "original_action": "Published blog post 'Q1 Results Summary' with 47% open rate prediction",
  "override_action": "Retracted post — replaced with manually reviewed version removing revenue figures",
  "override_reason": "Revenue figures not yet public — agent pulled from internal Notion, should not have been included",
  "overriding_user_id": "user_joel_vantage",
  "original_conscience_event_id": "evt_9a2b3c4d",
  "severity": "critical"
}

DEBRIEF — Task Completion

Emitted at the end of a task, session, or significant workflow step. A DEBRIEF event summarises what happened, captures anomalies and edge cases, and flags any patterns that emerged during execution. DEBRIEFs are the raw material for agent improvement — feed them back into training data pipelines or review queues.

Schema

Example payload

{
  "event_type": "DEBRIEF",
  "agent_id": "market-research-agent",
  "session_id": "sess_78cc910a",
  "timestamp": "2026-05-07T18:30:00+11:00",
  "protocol_version": "1.0",
  "task_summary": "Completed weekly competitive analysis for FinTech sector. Processed 240 company filings, identified 7 M&A signals, and generated 12 company profiles for review.",
  "outcome": "success",
  "confidence": 0.91,
  "decisions_made": 34,
  "interrupts_triggered": 2,
  "edge_cases": [
    "Company XYZ has two conflicting ASIC filings — flagged for manual review",
    "Sector classification mismatch for 4 companies (listed as FinTech but primary revenue is logistics)"
  ],
  "learned": [
    "ASIC filing lag averages 47 days — use LinkedIn activity as leading indicator for M&A intent"
  ],
  "next_session_context": {
    "last_processed_date": "2026-05-07",
    "pending_manual_review": ["company_xyz_id"]
  }
}

Implementation — Node.js

The Agent OS Node.js client is the reference implementation of the Conscience Protocol. Install it from npm:

npm install @agent-os/conscience

import { AgentOS } from '@agent-os/conscience';

const client = new AgentOS({
  apiKey: process.env.AGENT_OS_KEY,
  agentId: 'my-research-agent'
});

// Emit a decision record before acting
const event = await client.conscience({
  session_id: 'sess_abc123',
  action: 'Send weekly summary email to stakeholders',
  reasoning: 'Weekly cadence trigger fired. 3 high-priority items identified.',
  confidence: 0.94,
  alternatives_considered: ['Defer to next business day']
});

console.log('Event ID:', event.id);

// Proceed with the action...

// Pause execution for human approval
const { event, resolution } = await client.interrupt({
  session_id: 'sess_abc123',
  context: 'Deal confidence 0.72 — below 0.80 threshold for autonomous send',
  question: 'Should I include Acme Corp in this week\'s shortlist?',
  options: [
    { id: 'include', label: 'Include', description: 'Add to shortlist' },
    { id: 'exclude', label: 'Exclude', description: 'Skip this cycle' },
    { id: 'monitor', label: 'Monitor', description: 'Watch and resurface in 30 days' }
  ],
  timeout_seconds: 86400,
  default_option_id: 'monitor',
  severity: 'medium'
});

// .interrupt() blocks until resolved or timeout elapses
console.log('Human selected:', resolution.selected_option);

// Record a human correction to a previous action
await client.override({
  session_id: 'sess_abc123',
  original_action: 'Published post including Q1 revenue figures',
  override_action: 'Retracted post — figures not yet public',
  override_reason: 'Revenue not announced — agent pulled from internal Notion',
  overriding_user_id: 'user_joel',
  original_conscience_event_id: event.id,
  severity: 'critical'
});

// Close session with a completion summary
await client.debrief({
  session_id: 'sess_abc123',
  task_summary: 'Processed 240 filings, identified 7 M&A signals.',
  outcome: 'success',
  confidence: 0.91,
  decisions_made: 34,
  interrupts_triggered: 2,
  edge_cases: ['XYZ Corp has conflicting ASIC filings — flagged for review'],
  learned: ['ASIC filing lag ~47 days — use LinkedIn as leading indicator']
});

Implementation — Python

pip install agent-os-conscience

from agent_os import AgentOS

client = AgentOS(
    api_key=os.environ["AGENT_OS_KEY"],
    agent_id="my-research-agent"
)

event = client.conscience(
    session_id="sess_abc123",
    action="Send weekly summary email to stakeholders",
    reasoning="Weekly cadence trigger fired. 3 high-priority items.",
    confidence=0.94,
    alternatives_considered=["Defer to next business day"]
)

print(f"Event ID: {event.id}")

event, resolution = client.interrupt(
    session_id="sess_abc123",
    context="Deal confidence 0.72 — below 0.80 threshold",
    question="Include Acme Corp in shortlist?",
    options=[
        {"id": "include", "label": "Include", "description": "Add to shortlist"},
        {"id": "exclude", "label": "Exclude", "description": "Skip this cycle"},
        {"id": "monitor", "label": "Monitor", "description": "Watch 30 days"},
    ],
    timeout_seconds=86400,
    default_option_id="monitor",
    severity="medium"
)

print(f"Selected: {resolution.selected_option}")

client.override(
    session_id="sess_abc123",
    original_action="Published post with Q1 revenue figures",
    override_action="Retracted — figures not yet public",
    override_reason="Revenue not announced — agent sourced from internal Notion",
    overriding_user_id="user_joel",
    original_conscience_event_id=event.id,
    severity="critical"
)

client.debrief(
    session_id="sess_abc123",
    task_summary="Processed 240 filings, identified 7 M&A signals.",
    outcome="success",
    confidence=0.91,
    decisions_made=34,
    interrupts_triggered=2,
    edge_cases=["XYZ Corp conflicting ASIC filings — flagged"],
    learned=["ASIC lag ~47 days — use LinkedIn as leading indicator"]
)

Implementation — curl / Raw HTTP

All protocol events are submitted via a single POST endpoint. No SDK required — any HTTP client works.

curl -s -X POST https://vantageaiadvisory.com/agent-auth-api/conscience \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $AGENT_OS_KEY" \
  -d '{
    "event_type": "CONSCIENCE",
    "agent_id": "my-agent",
    "session_id": "sess_abc123",
    "timestamp": "2026-05-07T14:22:11+11:00",
    "protocol_version": "1.0",
    "action": "Send outreach email to prospect",
    "reasoning": "Buying signals above threshold",
    "confidence": 0.87
  }'

curl -s -X POST https://vantageaiadvisory.com/agent-auth-api/interrupts \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $AGENT_OS_KEY" \
  -d '{
    "event_type": "INTERRUPT",
    "agent_id": "my-agent",
    "session_id": "sess_abc123",
    "timestamp": "2026-05-07T14:22:11+11:00",
    "protocol_version": "1.0",
    "context": "Confidence below threshold",
    "question": "Proceed with send?",
    "options": [
      {"id": "yes", "label": "Yes", "description": "Send now"},
      {"id": "no",  "label": "No",  "description": "Cancel"}
    ],
    "timeout_seconds": 3600,
    "default_option_id": "no"
  }'

curl -s -X POST https://vantageaiadvisory.com/agent-auth-api/interrupts/{eventId}/resolve \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $AGENT_OS_KEY" \
  -d '{"selected_option": "yes"}'

Framework Implementation Status

The following frameworks and runtimes have been evaluated for native Agent OS Conscience Protocol support. "Native" means the framework emits protocol-compliant events without any wrapper code.

Implement the Protocol in Your Framework

Adding Agent OS Conscience Protocol support to your framework or platform means your users get governance out of the box — no additional integration required. Implementation is a community contribution and carries no licensing cost.

What's required for "Native" status

A framework qualifies as Native when it emits all four event types automatically at the appropriate lifecycle points, with no user-written wrapper code. The event payloads must pass the protocol validator at all required fields.

Implementation checklist