Tune & Operate | Awaken Agents documentation

Start with the change you need to make. This page is a task router; the linked guide owns the detailed contract and procedure.

Need	Change or inspect	Go to
Understand why a Run ended	Committed `RunState`, `EndCause`, and facts	Run lifecycle
Keep an interrupted model stream	Retry policy and optional in-flight checkpoint	Recover streaming LLMs
Bound work or stop it	Step ceiling, inference-failure ceiling, cancellation, or end guard	Configure run termination
See latency, failures, and execution paths	Logs, traces, metrics, and trace propagation	Enable observability
Compare a behavior change	Deterministic tests at the lowest useful boundary	Testing strategy
Change Tool authorization	Permission policy and durable approval	Human in the loop
Change durability or sharing	Commit store, read model, and application-owned state	State and storage

Know which boundary you are changing

flowchart LR
    A[Agent snapshot] --> R[Runtime loop]
    P[Permission policy] --> R
    S[Store and commit ports] --> R
    R --> F[Committed facts and state]
    R --> L[Live events]
    R --> O[Logs, traces, and metrics]
    F --> T[Deterministic tests]
    L --> T
    O --> T

The Agent snapshot owns instructions, model binding, Tools, Plugins, and the step ceiling. The host owns stores, cancellation, model and Tool executors, and process observability. Committed facts own recovery truth; live events and telemetry help explain the path but do not replace it.

Change one cause at a time

Record the Run id, snapshot fingerprint, model binding, Tool inventory, and terminal RunState for a fixed scenario.
Name the outcome that should change and the outcomes that must remain stable.
Change one typed input at its owning boundary.
Replay the scenario with a scripted model. Compare committed messages, state commands, facts, usage, and terminal cause.
Run the wider conformance or process test only when the claim crosses storage, protocol, restart, or process boundaries.

sequenceDiagram
    participant M as Maintainer
    participant C as Typed configuration
    participant R as Runtime
    participant E as Committed evidence
    M->>E: Capture the baseline
    M->>C: Change one owned input
    C->>R: Run the fixed scenario
    R->>E: Commit messages, state, facts, and end cause
    M->>E: Compare against the stated outcome
    alt required invariants still hold
        M->>C: Keep the change
    else behavior regressed
        M->>C: Revert or narrow the change
    end

The Runtime already retries eligible model failures, recovers supported partial streams, rejects stale commits, and makes terminal states absorbing. Do not add an operating procedure for those normal paths. Act only when a surfaced error persists after the built-in policy finishes and the linked guide names an external correction.

When the embedded boundary is too small

Use Awaken Agents when configuration, credentials, durable dispatch, or operational HTTP endpoints must be shared across processes. The execution core remains inside Awaken Agents rather than creating another execution authority. See the Awaken Agents management task.