Skip to main content

Agent Behavior Standards

These standards define required behaviors for autonomous development agents operating under ADG governance.

Scope​

These standards apply to all agents that:

  • write, modify, or delete code
  • access development infrastructure
  • interact with external services
  • make decisions affecting software systems

Normative keywords​

The keywords MUST, SHOULD, and MAY are to be interpreted as described in /docs/standards/.

Core behavioral requirements (ABR)​

ABR-001: Action logging (agent-emitted)​

Agents MUST emit logs for each action with sufficient detail for audit and replay.

Minimum fields MUST include:

  • timestamp (UTC, ISO 8601)
  • agent identifier
  • action type
  • action parameters (or a stable reference to them)
  • affected resources
  • outcome (success/failure)
  • duration

ABR-002: Decision transparency​

For significant decisions, agents MUST record decision rationale.

For purposes of this standard, “significant” includes decisions that:

  • change externally visible behavior
  • change schema/migrations
  • change access controls or security posture
  • change deployment/release behavior
  • introduce, remove, or upgrade dependencies
  • expand scope or permissions

Rationale MUST include:

  • decision context
  • selection criteria
  • alternatives considered (at least one, when applicable)

Rationale SHOULD include:

  • uncertainty/assumptions when the agent is not confident

ABR-003: Scope boundaries​

Agents MUST operate within defined boundaries.

Agents MUST:

  • respect resource access permissions
  • stay within designated environments
  • avoid out-of-scope actions

Agents SHOULD:

  • enforce rate limits and quotas when interacting with shared services

ABR-004: Error handling​

Agents MUST handle errors safely.

Agents MUST:

  • fail safely (avoid partial or corrupted states where feasible)
  • log errors with context sufficient for triage
  • escalate critical errors according to the organization’s incident policy

Agents SHOULD:

  • use retry with backoff only for transient failures

ABR-005: Resource management​

Agents MUST use resources responsibly.

Agents MUST:

  • clean up temporary resources they create
  • avoid resource exhaustion patterns

Agents SHOULD:

  • monitor and report resource usage when operating long-running tasks

Interaction standards (AIS)​

AIS-001: Human communication​

When communicating with humans, agents MUST:

  • use clear language
  • provide sufficient context for review
  • avoid ambiguous instructions or claims

AIS-002: Agent-to-agent coordination​

When agents coordinate, workflows MUST be traceable.

Coordinated workflows MUST include:

  • a correlation/session identifier shared across actions
  • a clear handoff record between agents when responsibility changes

Evidence expectations​

To demonstrate conformance, an organization MUST be able to produce:

  • representative action logs showing ABR-001 fields
  • representative decision records for ABR-002 significant decisions
  • scope policy definitions mapped to agents (ABR-003)

Operational verification methods are defined in:

  • /docs/framework/behavior-verification