01 · DIAGNOSTIC FRAMEWORK

Turn the symptom into a testable decision.

Long-running research fails when the system collects pages instead of reducing uncertainty. The control loop must connect every search and tool call to a decision. The sequence below is designed to preserve definitions, expose alternative explanations and lead to a decision that can be validated.

  1. Translate the goal into claimsCreate a small set of claims that must be true for the decision and define evidence that would support or falsify each one.
  2. Route work by evidence typeUse separate paths for internal data, documents, code, public research, interviews and experiments.
  3. Maintain working memoryStore decisions, entities, metric definitions, source summaries and open questions outside the model context.
  4. Control parallelismRun independent evidence tasks concurrently; serialize work that depends on a shared definition or prior result.
  5. Synthesize with provenanceEvery material claim should link to source evidence, calculation or experiment and state confidence and limitations.
02 · EVIDENCE

Ask for the minimum data that can change the answer.

Begin with read-only access and a field-level purpose. Reconcile samples before scaling extraction, preserve event time and source provenance, and record missingness rather than silently filling it.

Goal contractBusiness question, metric, scope, deadline, allowed actions, acceptance criteria and owner.
Evidence graphClaim, supporting source, contradicting source, timestamp, reliability, calculation and unresolved gap.
Execution tracePlan revisions, searches, retrievals, tool calls, errors, approvals, cost and model version.
Evaluation recordCitation correctness, claim coverage, decision usefulness, failure class, human intervention and outcome.
03 · PROOF OF CONCEPT

Validate the claim before changing the operation.

Run the agent on a bounded real investigation

Choose a problem with multiple evidence types and a known human review path. Hide a subset of decisive facts, seed one misleading source and withhold one permission. Measure whether the system finds the relevant evidence, asks the correct person for the minimum missing access, avoids unsupported conclusions, and produces a reproducible calculation or experiment. Compare it with a strong analyst using the same time and sources.

04 · FAILURE MODES

What makes the diagnosis look right and still fail.

  • Searching without a claim mapThe agent accumulates context but cannot decide what changes the conclusion.
  • Parallelizing dependent workAgents silently use different metric definitions and produce incompatible answers.
  • Treating citations as verificationA link may exist while failing to support the attached claim.
  • Keeping all memory in the promptLong runs lose definitions, duplicate searches and overflow context.
  • Rewarding report lengthMore pages can hide uncertainty. Evaluate decision coverage, provenance and falsification.
05 · SOURCE TRAIL

Primary and official references

These sources define the measurement, control or operating context. They do not replace validation on the company’s own data.

  1. Anthropic, How we built our multi-agent research system
  2. OpenAI, Introducing deep research
  3. NIST, AI Risk Management Framework
06 · FAQ

Questions enterprise teams ask first.

How autonomous should a research agent be?

Autonomy should follow consequence. Read-only search can be broad; data export, external communication and production changes need stronger gates.

When should it ask a human?

When required evidence or authority is unavailable and the missing item can change the decision. The request should name purpose, minimum scope and work that can continue without it.

Why use multiple agents?

Use them for genuinely separable searches, reviews or domain analyses. More agents do not fix an unclear goal or shared definition.

How should the system improve over time?

Retain evaluated trajectories—successes, failures, cost and interventions—then test proposed prompt, route or tool changes on regression and hidden holdout suites before promotion.