Core Concepts

Understand how Agent Flywheel captures runtime evidence, generates tasks, grades outcomes, trains candidates, and promotes approved agent behavior.

Agent Flywheel architecture

Agent or runtime SDK/plugin
        |
        | outbound runtime evidence and workflow state
        v
Mission Control [server-side work]
        |
        +--> Generate tasks
        +--> Grade outcomes
        +--> Train candidate
        +--> Compare current versus candidate
        +--> Approve promotion

Mission Control handles task generation, grading, training, promotion, and candidate-versus-current evaluation on the server. The local SDK or plugin connects outbound to Mission Control and does not reproduce those server-side operations.

Agent Flywheel lifecycle

Capture runtime evidence

The runtime SDK or plugin captures agent activity and tool evidence. Python integrations can provide explicit ground truth with record_tool_call().

Generate tasks

Mission Control uses captured data to generate tasks for the flywheel.

Grade outcomes

Mission Control grades task results from the available evidence. Outcomes are ok, failed, or unknown.

Train a candidate

Mission Control trains a candidate from eligible data.

Evaluate and approve

Mission Control compares candidate behavior with the current version. Promotion requires approval after the comparison meets your requirements.

Runtime and operator responsibilities

Responsibility	Runtime integration	Operator skill
Primary role	Produces runtime evidence.	Manages the Mission Control workflow.
Data source	Uses the runtime SDK or plugin to capture activity and tool evidence.	Uses eligible OTEL trace exports or data produced by the SDK.
Training and promotion	Does not reproduce server-side workflow operations.	Manages server-side training, grading, promotion, and related workflow operations.
Local training	Depends on the runtime integration mode.	Does not perform local training.

Runtime method reference

API or mode	Behavior	Boundary
`serve()`	Runs training rollouts and supports the rollout process used to produce training data and candidate behavior.	Claims rollouts for the training workflow.
`attach()`	Captures an existing production process without claiming rollouts.	Captures production activity but does not provide training rollouts.
`record_tool_call()`	Supplies explicit ground truth when the application knows the actual result of a tool operation.	Records the application's known verdict for a tool operation.

Capture capabilities depend on the runtime SDK or plugin you use.

Outcome statuses

Status	Meaning
`ok`	Observed success
`failed`	Observed failure
`unknown`	No trustworthy verdict

An unknown outcome is never a success. Preserve it when the available evidence does not support a trustworthy success or failure verdict.

Was this page helpful?