Skip to content

Agent lifecycle⚓︎

Every agent in ContextAgent follows the same lifecycle regardless of the underlying LLM provider. Understanding the lifecycle helps you inject custom logic and diagnose misbehaving runs.

1. Hydrate⚓︎

  • Load the agent's prompt template and defaults from the context core.
  • Merge run-specific state such as the user query, datasets, and tool outputs.
  • Resolve model + provider, picking overrides if supplied by the pipeline.

2. Plan⚓︎

Agents can request plan fragments before calling the LLM. Use this phase to:

  • Expand tasks into bulletproof checklists.
  • Benchmark available tools and select the minimal set required.
  • Emit guardrail prompts or self-check instructions.

Enable planning via the pipeline config (plan: true) and inspect traces in the observability UI.

3. Execute⚓︎

sequenceDiagram
  participant Context as Context Core
  participant Agent as Agent
  participant LLM as Model Provider
  participant Tools as Tool Bus
  Context->>Agent: Hydrated context package
  Agent->>LLM: Prompt + state
  LLM-->>Agent: Response chunk(s)
  Agent->>Tools: Optional tool call
  Tools-->>Agent: Tool result
  Agent->>Context: Final message + artifacts

During execution the agent can stream partial tokens back to the context core. Tool calls suspend the LLM stream until results arrive.

4. Reflect⚓︎

After the LLM returns, the agent optionally performs self-critique:

Self-reflection modules

  • Checklists — ensure required sections are present.
  • Classifiers — rate quality or completeness.
  • Automatic retries — re-run with additional instructions if metrics fall below thresholds.

Reflections are stored as artifacts so the pipeline can branch based on quality signals.

5. Hand off⚓︎

The agent emits a structured message back to the context core. Pipelines decide whether to:

  • Transition to a downstream agent.
  • Persist artifacts.
  • Return the final output to the caller.

Next, learn how the context engine coordinates shared state.