01 / Scope
Turn intent into a working brief.
Define the question, permitted sources, deliverables, and review boundaries before delegating. Keep the objective available to every role.
Benefit: a shared definition of completion makes scope changes easier to identify and review.
Objective · boundaries · acceptance criteria
Compare approaches + inspect the example +
- Direct model session
- A direct chat relies on the supplied instructions and conversation. The operator maintains scope across separate sessions.
- Agent frameworks
- A team framework can encode these rules. The application still needs explicit task contracts and approval policies.
- Jianmu example / evidence
- Example: investigate a payment chain within a specified period; separate confirmed links from open questions.
Inspect the scope and gate example ↗02 / Index
Query the corpus. Focus the context.
Organize records for structured queries. Retrieve relevant rows and source references so analysis can follow the evidence without repeatedly loading the entire corpus.
Benefit: separate storage volume from prompt volume; inspect a large corpus through focused results.
Indexed records · selected results · source references
Compare approaches + inspect the example +
- Direct model session
- Pasting extracted data into a chat spends context on the material itself. A model connected to retrieval tools can also query data.
- Agent frameworks
- Retrieval is compatible with multi-agent frameworks; the data connectors, schema, and citation handling determine the result.
- Jianmu example / evidence
- Recorded synthetic case: 2,749,851 indexed records, 13 tables, and 1,255 query events. Ingestion time is excluded.
Open the recorded data ledger ↗03 / Delegate
Keep the goal central. Distribute the work.
The L1 controller carries the overall objective. L2 coordinators organize workstreams; L3 specialists investigate bounded questions and return focused results.
Benefit: separate responsibilities and workstreams make complex work easier to coordinate and review.
Task packages · specialist results · handoffs
Compare approaches + inspect the example +
- Direct model session
- In a single direct session, one conversation carries planning and execution. Parallel workers require an orchestration layer.
- Agent frameworks
- AutoGen supports teams and handoffs too. Compare role boundaries and delivered workflow, not agent counts alone.
- Jianmu example / evidence
- Recorded: 114 handoff events and 34 distinct seats across the run. These are not simultaneous worker counts.
Inspect the coordination evidence ↗04 / Execute
Use reasoning where it adds value.
Use queries and repeatable tools for filtering, aggregation, and data checks. Give selected outputs to specialists for interpretation and further investigation.
Benefit: repeatable calculations stay inspectable; models spend their context on interpretation and decisions.
Tool results · calculations · analysis
Compare approaches + inspect the example +
- Direct model session
- A tool-enabled model can run code and queries. A direct text-only workflow requires the operator to supply those results.
- Agent frameworks
- Tool execution is common in agent systems. Compare how results are routed, logged, and reused across the complete task.
- Jianmu example / evidence
- Real-data replay: 216 comparable SQL records matched archived results, covering 110 distinct query/run combinations. This checks replay consistency.
Read the real-data verification ↗05 / Review
Challenge the finding. Inspect its basis.
Separate analysis from review. Check source references and alternative explanations; preserve rejected findings and unresolved questions alongside accepted results.
Benefit: expose weak claims before they become a final deliverable, with a traceable review decision.
Reviewed findings · corrections · open questions
Compare approaches + inspect the example +
- Direct model session
- Asking a model to check its own answer adds a review pass but does not establish independent correctness.
- Agent frameworks
- AutoGen also demonstrates a critic role. Compare source checks and evaluation outcomes, not simply the presence of a second agent.
- Jianmu example / evidence
- 38 proposals: 13 verified, 2 hypotheses, 23 rejected. A separate real-data audit also found unresolved citations; review labels are not accuracy scores.
Inspect review results and limits ↗06 / Continue
Carry forward state, not just a conversation.
Retain the objective, established findings, references, and next actions in a handover. Keep open issues visible when a new session takes over.
Benefit: reduce reconstruction work and make long-running investigations easier to resume without losing their direction.
Working state · handover · next actions
Compare approaches + inspect the example +
- Direct model session
- Starting a fresh chat requires relevant state to be supplied again, unless the surrounding application persists it.
- Agent frameworks
- LangGraph provides checkpoint persistence. Compare which domain findings and sources survive a handoff, alongside execution state.
- Jianmu example / evidence
- Illustrative budget: 1.4M controller-facing tokens, an 800K usable window, and 2 context segments. This is a planning calculation, not a measured run.
Explore the continuity calculation ↗