Research & intelligence workflows

Retrieval order the harness enforces

For knowledge tasks the system prompt defaults to: memory_query or recall_relevant → vault_search / vault_read → web_search and parallel web_fetch. That prevents skipping your own notes when answering “what did we decide about X?”

web_fetch uses readability extraction when AGENT_WEB_READABILITY is on, with timeouts and bot-wall retries. There is no separate web_research tool — the model composes search plus multiple fetches, which maps well to “gather five sources, compare claims.”

Parallelism without losing control

• spawn_agent — fork a child harness with a focused prompt (e.g. “only fetch SEC filings”).
• wait_for_agents / list_agents — coordinate batches.
• share_agent_context / read_agent_context — SharedMemoryBus keys for handoffs.
• dispatch_graph — intra-round DAG when tools depend on each other.
• synthesis_run — cross-domain synthesis sub-agent when angles must merge.

Dynamic workflows for multi-phase reports

plan_workflow builds a WorkflowSpec: phases, tasks per phase, optional adversarial review, verify gates (run_tests, run_lint, or critic). run_workflow executes waves of forkChild() agents, stores every raw result in .agent_workflows/<runId>/ (BM25-queryable via query_workflow), and injects only distilled phase summaries into the parent — so the parent context holds an executive report, not fifty full web pages.

Intent inference sets workflowSuitable when the user task fans into many independent sub-tasks (audits, multi-source research, cross-checked comparisons). The harness can auto-activate the workflow family so the model does not fall back to ad-hoc spawn loops.

Outputs researchers actually keep

Long briefs belong in vault_write with [[wikilinks]] and typed frontmatter (note, fact, episode). Short durable claims use remember({ scope: "workspace" }). extract_structured pulls JSON from messy HTML. The document engine (doc_* tools) can compile slides or PDFs when AGENT_DOC_ENGINE is on.

Large fetch bodies distill to .agent_artifacts/ pointers in context — the model re-reads via read_artifact instead of repeating megabytes each round.

Example research prompt

“Research competitors for X. Phase 1: official docs and pricing pages. Phase 2: engineering blog posts and GitHub issues. Phase 3: synthesize risks and opportunities; vault_write an executive summary with links; remember three atomic facts.”

Activate web + memory_advanced families, set AGENT_EFFORT=high for thorough deliverables, and optionally enable workflows for phase boundaries.