One-Line Split
ANATOP
Governed state formation. It governs how raw artifacts become admitted, bounded, deterministic, agent-usable state.
SSEJ
Governed decision reconstruction. It governs how a bounded run becomes a reconstructable record of what was seen, selected, excluded, and justified.
Runtime Example Boundary
Anthropic's public documentation for Claude Managed Agents describes a managed runtime for long-running, asynchronous agent work and currently labels the capability beta. Sovrient references Anthropic here as a concrete runtime example because the public docs expose stable nouns such as environment, session, events, and stream.
This page does not claim Anthropic exclusivity, partner status, or a current shipping Anthropic integration.
Current Claim Boundary
- No live public managed-agent decision lane is published today.
- No shipping Anthropic integration is claimed today.
- No claim is made that provider-internal model cognition is directly observable.
- The current public claim is architectural compatibility with managed-agent runtimes that expose bounded session and event surfaces.
Five-Layer Reference Architecture
1. Declared Scope
Define mission/task ID, runtime family, environment, tool boundary, policy profile, and publication boundary before execution.
Representative artifact: session_scope.json
2. ANATOP State Formation
Capture, fingerprint, validate, and admit the input surface that the run is actually allowed to use.
Representative artifacts: input_surface.json, environment_profile.json, admission_receipt.json
3. Managed Runtime Execution
Record provider-exposed session metadata, ordered events, tool activity, and emitted runtime artifacts within the declared boundary.
Representative artifacts: session_record.json, session_events.ndjson, tool_activity.json
4. SSEJ Reconstruction
Build the bounded run record: what the agent saw, what it selected, what remained excluded, and what justified the result under declared policy.
Representative artifacts: ssej_record.json, decision_summary.json, exclusion_register.json
5. Attestation And Review
Bind the run into replayable evidence, then support review queues, human resolutions, and narrow trigger payloads where policy allows.
Representative artifacts: state_receipt.json, attestation.json, review_queue.json, review_resolution.json
Minimum Public Artifact Contract
manifest.jsonfor package identity, runtime family, scope, and source precedence.session_scope.jsonfor task, tools, policy, and publication boundary.input_surface.jsonandenvironment_profile.jsonfor admitted run state.session_record.jsonandsession_events.ndjsonfor provider-exposed execution traces.tool_activity.jsonandruntime_artifact_manifest.jsonfor observed actions and files.ssej_record.jsonfor the bounded decision record.state_receipt.json,attestation.json, andverification_handles.jsonfor replay and proof.
What This Future Lane Would Prove
- What run boundary was declared.
- What admitted input surface was available to the bounded run.
- What provider-exposed events, tool uses, and artifacts were observed.
- What SSEJ record was reconstructed from that observable run surface.
- What review and attestation state followed.
What It Would Not Prove
- Hidden provider-internal reasoning or chain-of-thought state.
- Total completeness beyond the runtime fields the provider actually exposes.
- Legal fault, liability, or full-world truth outside the bundle boundary.
Phased Build Sequence
Phase 1. Operator-Local Demo
Build one bounded local lane over a declared session family and prove the artifact contract and SSEJ reconstruction over provider-exposed events.
Phase 2. Public Example Bundle
Publish one bounded example run with declared scope, admitted input surface, session event log, SSEJ record, and attestation.
Phase 3. Review And Trigger Extension
Add review queue, resolution ledger, and narrow payload classes so the lane becomes operationally useful rather than only technically demonstrative.