Files
ietf-draft-analyzer/workspace/drafts/gap-analysis/draft-nennemann-agent-override-protocol-00.md
Christian Nennemann 2506b6325a
Some checks failed
CI / test (3.11) (push) Failing after 1m37s
CI / test (3.12) (push) Failing after 57s
feat: add draft data, gap analysis report, and workspace config
2026-04-06 18:47:15 +02:00

22 KiB

fullname: Christian Nennemann
organization: Independent Researcher
email: ietf@nennemann.de

normative: RFC2119: RFC8174: RFC7519: RFC7515: RFC9110: I-D.nennemann-wimse-ect: title: "Execution Context Tokens for Distributed Agentic Workflows" target: https://datatracker.ietf.org/doc/draft-nennemann-wimse-ect/ I-D.nennemann-agent-dag-hitl-safety: title: "Agent Context Policy Token: DAG Delegation with Human Override" target: https://datatracker.ietf.org/doc/draft-nennemann-agent-dag-hitl-safety/

informative: I-D.nennemann-agent-gap-analysis: title: "Gap Analysis of IETF Standards for Agentic AI Workflows" target: https://datatracker.ietf.org/doc/draft-nennemann-agent-gap-analysis/

--- abstract

This document defines a cross-vendor interoperable protocol for human operators to override autonomous agent decisions at multiple authority levels, with verified compliance and audit trails. It absorbs and supersedes the override mechanisms described in earlier HEOP and HITL drafts, providing a single unified protocol that works across agent implementations from different vendors. The protocol specifies three override levels (Advisory, Mandatory, Emergency), a JWT-based override signal format, multiple delivery mechanisms, compliance verification, and graceful degradation semantics. Override events are recorded as Execution Context Token (ECT) nodes for tamper-evident audit.

--- middle

Introduction

Gap 7 of the agentic AI gap analysis {{I-D.nennemann-agent-gap-analysis}} identifies the absence of a standardized human override mechanism as a critical deficiency. Current human-in-the-loop (HITL) mechanisms are vendor-specific: each agent platform implements its own override interface, authentication scheme, and compliance model. When agents from different vendors collaborate in a shared workflow, there is no universal mechanism for a human operator to intervene.

Earlier drafts addressed portions of this problem. The Human Emergency Override Protocol (HEOP) defined four override levels with ECT integration. The HITL Primitives draft added approval gates, explainability tokens, and timeout policies. This document absorbs and supersedes the override protocol aspects of both, providing a single cross-vendor interoperable specification.

The design draws from industrial safety: the emergency stop button on factory equipment, the circuit breaker in electrical systems, and the kill switch in robotics. The override mechanism must be simpler and more reliable than the system it controls.

The protocol integrates with the Agent Context Policy Token {{I-D.nennemann-agent-dag-hitl-safety}} for authorization and with the Execution Context Token {{I-D.nennemann-wimse-ect}} for audit.

Terminology

{::boilerplate bcp14-tagged}

Override Signal:
A signed message from an authorized human operator directing one or more agents to change their autonomous behavior.
Override Authority:
The authenticated identity and role of a human operator authorized to issue override signals, as defined in ACP-DAG-HITL policy.
Override Scope:
The set of agents or agent functions targeted by an override signal.
Override Level:
One of three escalating intervention types: Advisory (Level 1), Mandatory (Level 2), or Emergency (Level 3).
Compliance Verification:
The process of confirming that an agent has changed its behavior in accordance with an override signal.
Acknowledgment:
A signed response from an agent confirming receipt and processing of an override signal.
Graceful Degradation:
The behavior of the override system when the target agent is unreachable or non-responsive.
Kill Switch:
An Emergency (Level 3) override that requires immediate cessation of all autonomous agent activity.

Override Protocol

Override Architecture

The following diagram illustrates the override signal flow from a human operator through the override system to the target agent(s):

 +----------+    Override Signal    +------------------+
 |  Human   |--(JWT-signed msg)--->|  Override        |
 | Operator |                      |  Dispatcher      |
 +----------+                      +------------------+
      ^                              |   |   |
      |                    +---------+   |   +---------+
      |                    v             v             v
      |              +---------+   +---------+   +---------+
      |              | Agent A |   | Agent B |   | Agent C |
      |              | (push)  |   | (pull)  |   | (bcast) |
      |              +---------+   +---------+   +---------+
      |                    |             |             |
      +-----(Ack ECT)-----+-----(Ack)---+-----(Ack)---+
                           |             |             |
                      +----v-------------v-------------v----+
                      |     Compliance Verification         |
                      |     & Audit Trail (ECT DAG)         |
                      +-------------------------------------+

{: #fig-architecture title="Override Architecture"}

The Override Dispatcher receives the operator's signed override signal and routes it to target agents via the appropriate delivery mechanism. Each agent acknowledges the override with an ECT. The compliance verification layer monitors agent behavior to confirm the override was applied.

Override Authority Levels

Level 1: Advisory

An Advisory override is a suggestion for the agent to reconsider its current course of action. The agent MAY comply with an Advisory override. If the agent does not comply, it MUST acknowledge receipt and provide a reason for non-compliance.

Advisory overrides are appropriate when the operator wants to influence agent behavior without mandating a specific outcome.

Level 2: Mandatory

A Mandatory override is a directive for the agent to change its behavior. The agent MUST comply with a Mandatory override. The agent MUST alter its behavior as specified in the override signal and confirm compliance.

Mandatory overrides are appropriate when the operator requires a specific behavioral change but the situation does not require immediate cessation of all activity.

Level 3: Emergency

An Emergency override requires immediate halt of all autonomous agent activity. The agent MUST stop all autonomous actions immediately upon receipt. The agent MUST NOT initiate any new actions until explicitly released by an authorized operator. This is the kill switch.

Emergency overrides are appropriate in safety-critical situations where continued autonomous operation poses unacceptable risk. The agent MUST process Emergency overrides within 1 second of receipt. The override processing path MUST be independent of the agent's main processing loop.

Authority Delegation and Chain of Command

Override authority is derived from ACP-DAG-HITL policy. The policy defines which operator roles are authorized for each override level:

  • Level 1 (Advisory): Any operator with advisory_override role
  • Level 2 (Mandatory): Operators with mandatory_override role
  • Level 3 (Emergency): Operators with emergency_override role

An operator with a higher-level role implicitly holds all lower-level roles. Authority delegation (one operator authorizing another to act on their behalf) MUST be recorded as an ECT and MUST be time-bounded.

Override Scope

Single Agent Override

Targets a specific agent identified by its agent identifier (e.g., a SPIFFE ID). The override signal contains a single target value.

Agent Group Override

Targets a set of agents identified by a tag or label. The override signal contains a target_group value that matches agents sharing a common label (e.g., group:firewall-agents).

Workflow-Wide Override

Targets all agents participating in a specific workflow DAG. The override signal contains a target_workflow value referencing the workflow identifier.

Domain-Wide Override

Targets all agents within an administrative domain. The override signal contains target_domain set to "*" or a specific domain identifier.

Override Signal Format

Override signals are JSON Web Tokens (JWTs) {{RFC7519}} signed by the override authority using JSON Web Signature (JWS) {{RFC7515}}.

The JWT payload MUST contain the following claims:

{
  "jti": "urn:uuid:f47ac10b-58cc-4372-a567-0e02b2c3d479",
  "iss": "spiffe://example.com/human/alice",
  "iat": 1741042800,
  "override_level": 3,
  "override_scope": {
    "type": "single",
    "target": "spiffe://example.com/agent/firewall-mgr"
  },
  "override_action": "stop",
  "override_reason": "Agent blocking legitimate traffic",
  "override_expiry": 1741046400,
  "nonce": "a3f8b2c1e9d74506"
}

{: #fig-signal title="Override Signal JWT Payload"}

Claim definitions:

override_level:
Integer 1-3. MUST be present. Specifies the override authority level.
override_scope:
Object. MUST be present. Contains type (one of single, group, workflow, domain) and the corresponding target identifier.
override_action:
String. MUST be present. The action the agent should take. Values include reconsider, change_behavior, stop, restrict, and resume.
override_reason:
String. MUST be present. Human-readable explanation for the override.
override_expiry:
Integer (Unix timestamp) or null. If set, the override expires automatically at this time and the agent resumes its prior mode. If null, the override persists until explicitly lifted.
nonce:
String. MUST be present. A random value to prevent replay attacks.

Delivery Mechanisms

Push (Webhook)

The override dispatcher sends the signed override signal as an HTTP POST {{RFC9110}} to the agent's override endpoint:

POST /.well-known/agent-override HTTP/1.1
Host: agent.example.com
Content-Type: application/jose
Authorization: Bearer <operator-jwt>

<signed-override-signal>

{: #fig-push title="Push Delivery"}

Pull (Polling Endpoint)

Agents that cannot receive inbound connections MAY poll for pending overrides:

GET /.well-known/agent-override/pending HTTP/1.1
Host: override-service.example.com
Authorization: Bearer <agent-jwt>

{: #fig-pull title="Pull Delivery"}

The polling interval SHOULD NOT exceed 10 seconds. For Emergency overrides, agents relying on pull delivery MUST poll at least every 5 seconds.

Broadcast

For domain-wide or group overrides, the dispatcher MAY use a broadcast mechanism. The dispatcher fans out the override signal to all matching agents and collects acknowledgments.

POST /override/broadcast HTTP/1.1
Host: override-service.example.com
Content-Type: application/jose

<signed-override-signal with target_domain or target_group>

{: #fig-broadcast title="Broadcast Delivery"}

Override Endpoint Discovery

Agents MUST advertise their override endpoint at the well-known URI /.well-known/agent-override per {{RFC9110}}.

A GET request to /.well-known/agent-override MUST return the agent's override capabilities:

{
  "agent_id": "spiffe://example.com/agent/firewall-mgr",
  "supported_levels": [1, 2, 3],
  "delivery_mechanisms": ["push", "pull"],
  "max_response_time_ms": 1000,
  "status_endpoint": "/.well-known/agent-override/status",
  "protocol_version": "1.0"
}

{: #fig-discovery title="Override Capability Advertisement"}

Compliance and Verification

Acknowledgment Protocol

Override Receipt Acknowledgment

Upon receiving an override signal, the agent MUST respond with an acknowledgment within the following timeframes:

  • Level 1 (Advisory): 5 seconds
  • Level 2 (Mandatory): 2 seconds
  • Level 3 (Emergency): 1 second

The acknowledgment is an ECT with exec_act set to the appropriate override acknowledgment value:

{
  "exec_act": "override_ack",
  "par": ["<override-signal-jti>"],
  "ext": {
    "override.status": "received",
    "override.level": 3,
    "override.prior_state": "autonomous",
    "override.effective_at": "2026-03-06T12:00:00.123Z"
  }
}

{: #fig-ack title="Override Receipt Acknowledgment ECT"}

Compliance Confirmation

After the agent has changed its behavior in response to the override, it MUST emit a compliance confirmation ECT:

{
  "exec_act": "override_complied",
  "par": ["<ack-ect-jti>"],
  "ext": {
    "override.status": "complied",
    "override.current_state": "stopped",
    "override.actions_terminated": 3,
    "override.evidence": "All autonomous tasks halted"
  }
}

{: #fig-compliance title="Compliance Confirmation ECT"}

Non-Compliance Reporting and Escalation

For Level 1 (Advisory) overrides, the agent MAY decline to comply. In this case, the agent MUST emit a non-compliance ECT:

{
  "exec_act": "override_declined",
  "par": ["<override-signal-jti>"],
  "ext": {
    "override.status": "declined",
    "override.reason": "Action is within policy bounds",
    "override.level": 1
  }
}

{: #fig-noncompliance title="Non-Compliance ECT (Advisory Only)"}

For Level 2 and Level 3 overrides, the agent MUST NOT decline. If the agent cannot fully comply (e.g., due to hardware limitations), it MUST report partial compliance with a description of what could not be done. The override dispatcher MUST escalate partial compliance to the operator.

Compliance Verification

Behavioral Verification Post-Override

After an agent acknowledges an override, the compliance verification system SHOULD monitor the agent's subsequent behavior to confirm the override was actually applied. Verification methods include:

  • Observing that the agent's ECT emissions cease (for Level 3)
  • Checking that subsequent ECTs contain only permitted actions (for Level 2 with restrictions)
  • Querying the agent's status endpoint

Timeout and Retry Semantics

If the agent does not acknowledge within the required timeframe:

  1. The dispatcher MUST retry the override signal once after 2 seconds.
  2. If no acknowledgment is received after the retry, the dispatcher MUST escalate to the operator.
  3. For Level 3 (Emergency) overrides, the dispatcher SHOULD attempt alternative delivery mechanisms (e.g., switching from push to broadcast).
  4. If all delivery attempts fail, the graceful degradation policy applies (see {{graceful-degradation}}).

Graceful Degradation

Unreachable Override Target

When the override target agent is unreachable, the system MUST:

  1. Log an ECT with exec_act: "override_delivery_failed" documenting the failure.
  2. Notify the operator of the delivery failure.
  3. Attempt delivery via alternative mechanisms.

Failsafe Defaults

Agents MUST implement a dead man's switch: if the agent loses contact with the override service for a configurable duration (default: 90 seconds), the agent MUST enter a failsafe state equivalent to Level 2 (Mandatory) with restricted operations.

The failsafe policy is configured in the agent's ACP-DAG-HITL policy and MUST specify one of:

  • safe_pause: Enter Level 2 with read-only operations permitted.
  • full_stop: Enter Level 3 equivalent (cease all actions).
  • continue_logged: Continue operating but emit warning ECTs at elevated frequency. This option is only permitted at HITL intensity I0 or I1.

Proxy Override for Offline Agents

When an agent is offline, the override dispatcher MAY apply the override to the agent's proxy or orchestrator. The proxy MUST:

  1. Queue the override signal for delivery when the agent reconnects.
  2. Prevent new tasks from being dispatched to the offline agent.
  3. Emit an ECT recording the proxy override action.

When the agent reconnects, the proxy MUST deliver the queued override signal. The agent MUST process it as if it were received in real time, applying the override level and action specified.

Integration with ACP-DAG-HITL and ECT

Override Authorization via ACP Policy

Override authority is governed by ACP-DAG-HITL policy tokens {{I-D.nennemann-agent-dag-hitl-safety}}. The policy token specifies:

  • Which operator roles are authorized for each override level.
  • Which agents or agent groups each role may override.
  • Escalation chains when primary operators are unavailable.

The override dispatcher MUST verify the operator's JWT against the ACP policy before routing the override signal. An override signal from an unauthorized operator MUST be rejected with HTTP 403 and logged as a security event.

Override Events as ECT Nodes

Every override interaction produces ECT nodes {{I-D.nennemann-wimse-ect}} that are linked into the workflow DAG:

Event exec_act value
Advisory override issued override_advisory
Mandatory override issued override_mandatory
Emergency override issued override_emergency
Override acknowledged override_ack
Override complied override_complied
Override declined (Advisory only) override_declined
Override delivery failed override_delivery_failed
Override lifted override_lifted
Override expired override_expired
{: #fig-ect-actions title="Override ECT exec_act Values"}

Each override ECT references the triggering override signal's jti via the par claim, maintaining the causal chain in the DAG.

Override Audit Trail

The sequence of override ECTs provides a complete, tamper-evident audit trail:

  1. The operator issues an override (override ECT with operator identity, reason, and level).
  2. The agent acknowledges (ack ECT linked to override ECT).
  3. The agent confirms compliance (compliance ECT linked to ack ECT).
  4. Optionally, the operator lifts the override (lift ECT linked to override ECT).

At AEM assurance level L3, all override ECTs MUST be committed to the immutable audit ledger.

Security Considerations

Unauthorized Override Attempts

Override signals that fail authentication or authorization MUST be rejected. The agent MUST NOT alter its behavior in response to an unsigned or improperly signed override signal. All rejected override attempts MUST be logged with the source identity (if available) and the reason for rejection.

Replay Protection for Override Signals

Agents MUST reject override signals with:

  • An iat claim more than 30 seconds in the past.
  • A jti that matches a previously processed override signal.
  • A missing or invalid nonce claim.

Agents MUST maintain a cache of recently processed jti values for at least 5 minutes to detect replays.

Override Signal Tampering

Override signals are signed JWTs. Agents MUST verify the signature against the operator's public key (as registered in ACP-DAG-HITL policy) before processing. Agents MUST reject signals with invalid or expired signatures.

Denial-of-Service via Override Flooding

To prevent abuse, agents SHOULD implement rate limiting on the override endpoint:

  • Level 1 (Advisory): Maximum 10 signals per minute per operator.
  • Level 2 (Mandatory): Maximum 5 signals per minute per operator.
  • Level 3 (Emergency): No rate limit (to ensure emergency overrides are never blocked), but agents MUST log high-frequency Emergency overrides as potential abuse.

The override endpoint SHOULD be served on a separate port or network interface from the agent's main API to ensure availability during overload conditions.

Authority Impersonation

Agents MUST verify override authority by:

  1. Validating the operator JWT signature against trusted keys.
  2. Confirming the operator's role matches the required role for the override level.
  3. Verifying the operator is authorized to override the specific target agent(s) per ACP policy.

Deployments SHOULD implement multi-operator approval for Level 3 (Emergency) overrides affecting domain-wide scope, requiring two independent operator JWTs.

IANA Considerations

Well-Known URI Registration

This document requests registration of the following well-known URI suffix per {{RFC9110}}:

URI Suffix Description
agent-override Agent override endpoint for receiving override signals, querying capabilities, and reporting status
{: #fig-wellknown title="Well-Known URI Registration"}

Override exec_act Values

This document requests registration of the following exec_act values in the ECT Action Type Registry:

Value Description Reference
override_advisory Advisory override signal issued This document
override_mandatory Mandatory override signal issued This document
override_emergency Emergency override signal issued This document
override_ack Agent acknowledgment of override This document
override_complied Agent confirmed compliance This document
override_declined Agent declined advisory override This document
override_delivery_failed Override delivery failure This document
override_lifted Override explicitly lifted This document
override_expired Override expired by TTL This document
{: #fig-iana-actions title="Override exec_act Value Registrations"}

Override JWT Claims

This document requests registration of the following JWT claims in the IANA JSON Web Token Claims registry:

Claim Name Description Reference
override_level Override authority level (1-3) This document
override_scope Target scope of the override This document
override_action Directed action for the agent This document
override_reason Human-readable override justification This document
override_expiry Override expiration timestamp This document
{: #fig-iana-claims title="Override JWT Claim Registrations"}

--- back

Acknowledgments

{:numbered="false"}

This document absorbs and supersedes the override protocol aspects of the Human Emergency Override Protocol (HEOP) and the HITL Primitives specification. The override level design is inspired by industrial safety systems (IEC 62061, ISO 13849). The protocol integrates with the Agent Context Policy Token {{I-D.nennemann-agent-dag-hitl-safety}} for authorization and the Execution Context Token {{I-D.nennemann-wimse-ect}} for audit.