Agent Atlas / Chapter 9

Trust Must Make Its Case

Trustworthy behavior is not a personality trait or product promise. It is an assurance case built from inspectable policy provenance, durable control state, bounded execution, traceable lineage, recoverable records, and explicit residual risk.

Learning objective

Use public evidence to answer: What evidence would justify an agent's authority before and after it acts?

Synthesis

What evidence would justify an agent's authority before and after it acts?

Trust is deserved when a system can state its operating claim, produce evidence that the claim held for this run, and show where evidence remains incomplete.
An assurance chain connects a bounded trust claim to policy provenance, runtime controls, action lineage, recovery evidence, and explicit residual risk.
From trust claim to residual risk A trust claim requires policy, runtime, lineage, and recovery evidence while unresolved risks remain visible. Open full-size diagram
Case file 09 / The permission that survived the handoff

Can the runtime prove which rules governed the child?

Consider a resumed thread that delegates work to a child. The assurance question is not whether the interface displays a reassuring approval label. It is whether the runtime can prove which permission profile, tool mode, policy layers, and parent identity governed the child when it acted.

Recent Codex evidence provides parts of that case: permission profiles persist through thread storage, resolved tool modes travel on TurnContext, and parent-thread lineage is added to turn metadata. These mechanisms improve inspectability, but they do not prove that every downstream action obeyed the recorded controls.

Latest newsroom receipts

What changed since the first Atlas draft

An assurance case is more than a receipt pile

A useful assurance case begins with a bounded claim: this run used these policies, exposed these tools, inherited this authority, and produced these actions. It connects the claim to implementation evidence and names what the evidence cannot establish.

Commit counts and generic source lists are not enough. The important evidence is relational: which policy won, which control reached the turn, which worker owned the action, and whether the record remained intact through resume, delegation, failure, and recovery.

Policy needs provenance and conflict behavior

Codex requirements-layer composition adds ordered precedence, domain-specific merging, provenance, additive deny-read behavior, and fail-closed handling for conflicting managed hook directories. Policy combination becomes an inspectable operation rather than an unexplained final configuration.

Cloud-managed config support preserves backend-supplied layer identities through diagnostics, debug output, hook attribution, and app-server surfaces. Provenance can explain where a rule came from without proving that every runtime consumer applied it.

The control envelope must survive movement

Agent work crosses boundaries constantly: turns resume, models switch, workers spawn, and execution moves between local and remote paths. Trust weakens whenever active control state is reconstructed from assumptions instead of carried as explicit runtime data.

Persisted permission profiles, ToolMode on TurnContext, and parent-thread plus subagent-kind lineage create a stronger chain of custody. They let an inspector ask whether the same control envelope survived storage, model selection, delegation, and later analysis.

Recovery is evidence, not absolution

Hermes undo rewinds by soft-deleting transcript rows for audit and rebuilding the next context from active history. Codex separately exposes parent lineage so subagent relationships remain distinguishable from copied-history forks.

Those mechanisms improve reconstructability, but they do not reverse every external effect. A transcript can be rewound while a file write, network request, or remote action remains. Assurance must separate recoverable conversation state from irreversible world state.

Synthesis map

The assurance case assembled

  1. Bounded claim

    known and limited authority

  2. Policy evidence

    layers, precedence, provenance

  3. Runtime evidence

    permissions and tool mode

  4. Action lineage

    worker, call, effect

  5. Recovery evidence

    rewind and reconstruction

  6. Residual risk

    gaps remain explicit

Failure modes

Where the contract breaks

Provenance theater

The runtime records where policy came from, but a consumer ignores the composed result.

Failure signalDiagnostics name a restrictive layer while observed behavior contradicts it.

Control-state split brain

Persisted metadata diverges from effective execution state.

Failure signalA resumed thread behaves differently without a recorded control transition.

Reversible transcript, irreversible world

Undo repairs conversation history while external effects remain active.

Failure signalReconstructed context omits actions still visible in connected systems.
Release tests

What the product must prove

  1. Show the effective policy layers, their precedence, and the provenance of winning values.
  2. Compare persisted permission, tool-mode, and lineage metadata with controls used for a resumed or delegated turn.
  3. Separate transcript recovery from external-effect recovery and identify actions requiring compensation.

Projects In This Lesson

24 projects

Recent Reporting For This Lesson

6 articles