Skip to content

Agent execution

Partial

Research, synthesis, assessment and clarification exist; arbitrary adaptive tools and child-agent graphs do not.

Code and tests

The worker currently executes a bounded sequence: load authorized context, obtain a structured research plan, gather sources, synthesize, assess and optionally refine. It can pause for clarification and reuse retained stages. It does not yet choose freely among arbitrary tools or child-agent graphs.

flowchart TD
  Context["Load scoped context"] --> Plan["Structured research plan"]
  Plan --> Ask["Clarification when requested"]
  Ask --> Resume["Persist reply and continue task"]
  Plan --> Research["Gather legal evidence"]
  Research --> Draft["Synthesize"]
  Draft --> Check["Validate evidence and assess gaps"]
  Check --> Refine["Bounded refinement if needed"]
  Refine --> Result["Publish result or limitations"]

The worker validates a research plan before using it, because model output cannot authorize arbitrary capabilities. It checks quotations against the selected source text and rejects an assessment that claims success while listing unresolved gaps. Clarification and completed stages are persisted, so reconnecting the browser does not restart the task.

For example, a generated answer quoting text that is absent from its selected source is rejected by evidence validation. A valid quote can still be legally irrelevant: semantic support and completeness depend on the assessment and need evaluated scenarios; they are not guaranteed by string matching.

Adaptive effort selection, targeted claim repair and general child-agent planning are targets. See Durable execution for the implemented command lifecycle and Workers & recovery for publication authority.