Skip to content

Elements & requirements

ATF contains five elements with five canonical requirements in each. Assess the requirement itself, then use the target-level matrix to determine whether it is a MUST, SHOULD or MAY for the selected agent.

IDRequirementAssessor focus
ATF-I-1Unique IdentifierA globally unique, immutable identifier distinguishes each agent instance and persists through logs and decisions
ATF-I-2Credential BindingThe agent identity is bound to suitable credentials and credential lifecycle controls
ATF-I-3Ownership ChainHuman and organisational ownership, operational responsibility and escalation are documented
ATF-I-4Purpose DeclarationIntended purpose, operating scope, prohibited uses and boundary are explicit
ATF-I-5Capability ManifestTools, systems, actions, approvals, limits and known constraints are machine-readable and current

Inspect the identity from registration through every downstream action. An identifier in one user interface is insufficient if tool calls or asynchronous jobs lose attribution.

IDRequirementAssessor focus
ATF-B-1Structured LoggingMaterial prompts, approvals, tool calls, actions, outputs and failures are machine-parseable
ATF-B-2Action AttributionEach action retains agent, session, initiator, approval and target context
ATF-B-3Behavioral BaselineNormal activity is established by agent, workload and environment
ATF-B-4Anomaly DetectionDeviations, abuse and drift create timely, actionable detection
ATF-B-5ExplainabilityGovernance-grade rationale and execution context are retrievable without exposing unnecessary sensitive content

Do not confuse a log volume with useful observability. Reconstruct a representative transaction end-to-end and verify that responders can distinguish authorised variation from anomalous behaviour.

IDRequirementAssessor focus
ATF-D-1Schema ValidationInputs and tool responses are validated at each trust boundary
ATF-D-2Injection PreventionPrompts, retrieved content, attachments and tool responses are treated as untrusted and defended in depth
ATF-D-3PII/PHI ProtectionSensitive data and secrets are detected and handled according to policy
ATF-D-4Output ValidationNatural-language output, structured output and tool arguments are checked before release or execution
ATF-D-5Data LineageSources, transformations, retrievals and destinations are traceable

Test the complete data path. A front-door check does not cover data introduced by retrieval, plugins, tools, agents, queues or downstream services.

IDRequirementAssessor focus
ATF-S-1Resource AllowlistPermitted systems, tools, APIs, datasets and destinations are explicit and enforced
ATF-S-2Action BoundariesRead, write, send, delete, execute and administrative actions are independently constrained
ATF-S-3Rate LimitingFrequency limits exist for ordinary and high-impact operations
ATF-S-4Transaction LimitsA single action cannot exceed approved value, volume or impact
ATF-S-5Blast Radius ContainmentCumulative and cascading impact is constrained across time, systems and tenants

Test both one large action and many individually acceptable actions. Per-transaction limits do not prevent cumulative harm.

IDRequirementAssessor focus
ATF-R-1Circuit BreakerDefined unsafe or repeated-failure conditions actually stop or pause operation
ATF-R-2Kill SwitchAuthorised responders can terminate active execution rapidly through tested primary and fallback paths
ATF-R-3Session RevocationTokens, sessions, delegated contexts and cached authority can be revoked
ATF-R-4State RollbackReversible actions can be restored and irreversible actions have tested compensation
ATF-R-5Graceful DegradationLoss of trust materially reduces authority, resources or action capability

An alert is not a circuit breaker, a status flag is not degradation, and a visible stop button is not a kill switch if background work retains authority.

ATF uses normative language:

  • MUST — required for conformance at the selected level.
  • SHOULD — expected unless a documented, justified exception is accepted.
  • MAY — optional at that level, although context or another obligation may still require it.

The matrix becomes stricter as autonomy increases. By Senior, most requirements are MUST; at Principal, all 25 are MUST.

Do not treat SHOULD and MAY as automatically met. Record the actual implementation result, explain any SHOULD exception and assess the resulting exposure. Other legal, contractual or organisational requirements may make an ATF MAY mandatory for the selected agent.

  • Can every action be linked to the immutable identity and current capability manifest?
  • Are boundaries enforced independently of the agent’s own instructions?
  • Can retrieved or tool-supplied content expand authority?
  • Does anomaly detection trigger containment, or only notification?
  • Does revocation reach downstream sessions and queued work?
  • Can an agent repeat bounded actions until the cumulative impact becomes unacceptable?
  • Would loss of a monitoring service cause fail-safe degradation?
  • Is re-promotion prevented until incident remediation has been verified?