agentic-research-operations

name	agentic-research-operations
description	Use when designing, running, or auditing multi-agent research operations: wave planning, task briefs, agent roles, tool boundaries, verification loops, failure recovery, evaluation, and portable runtime behavior across Codex, Claude, and other agents.
metadata	{"portable":true,"compatible_with":["claude-code","codex"]}

Agentic Research Operations

Use When

• Use when a project dispatches multiple research agents, verification agents, synthesis agents, or code agents.
• Use when building or changing orchestration, agent prompts, runtime portability, or evaluation gates.
• Use when a research workflow must be reproducible, auditable, and resilient to agent drift.

Do Not Use When

• A single local edit or fact-check needs no orchestration.
• The user explicitly asks not to delegate or parallelize.

Operations Intake Guidance

• Research goal, cohorts, source classes, outputs, risks, tools, time horizon, and verification standard.
• Current project context and any prior wave outputs.
• The hard evidence-discipline boilerplate for every sub-agent prompt.

Legacy Operations Method

1. Decide whether agents help: split only independent work with clear boundaries and verifiable outputs.
2. Define roles: researcher, verifier, gap analyst, red-team reviewer, synthesizer, renderer, or code worker.
3. Write briefs with goal, scope, out-of-scope, source classes, deliverable, storage path, and hard constraints.
4. Assign disjoint ownership: cohorts, files, source classes, or verification questions.
5. Require structured outputs: findings, sources, gaps, confidence, and next actions.
6. Verify before merge: source liveness, quote checks, contradiction scan, and claim graph update.
7. Keep cross-cohort synthesis with the orchestrator unless a task is explicitly a bounded synthesis subproblem.
8. Record operational lessons in project status or evidence audit.

Quality Standards

• Delegated tasks are independent, bounded, and material to the outcome.
• Agents receive enough context to avoid duplicate work but not so much that verification independence is lost.
• Every agent output has a merge criterion and rejection criterion.
• Tool access is least-privilege for the task.
• Verification is a separate pass, not trust in a polished answer.

Anti-Patterns

• Delegating the critical path when the orchestrator is blocked on it.
• Multiple agents researching the same broad question without cohort boundaries.
• Asking a research agent to synthesize across cohorts it cannot see.
• Accepting sub-agent citations without verification.
• Hiding failures instead of logging them.

Legacy Operations Deliverables

• Agent wave plan.
• Standard brief.
• Verification checklist.
• Merge/reject decision.
• Operations lesson log.

Evidence Produced

Category	Artifact	Format	Example
Operability	Wave plan	Markdown table	Agent, cohort, scope, output, verification
Release evidence	Merge record	Markdown table	Agent output accepted/rejected with reason

References

• Load references/agent-brief-template.md before dispatching agents.
• Load references/agent-evaluation-loop.md when accepting, rejecting, or improving agent outputs.

Companion Skills

Inputs

Input	Source/provider	If absent
Research plan, cohort boundaries, evidence clause	Orchestrator and project files	Stop dispatch until boundaries and evidence rules exist
Runtime capabilities and permissions	Host runtime	Use the least-capable read-only route and mark unavailable work

Capability Contract

Agents default to read-only research. File mutation, external messages, publication, destructive action, spending, and certification require explicit authority and bounded ownership.

Degraded Mode

If delegation or tools are unavailable, run waves sequentially and return partial evidence with unassessed tasks and recovery checkpoints named.

Decision Rules

Choice	Action	Failure/risk avoided
Workstreams are independent	Delegate with non-overlapping ownership	Merge conflict
Shared file is involved	Keep with orchestrator	Concurrent overwrite
Agent violates evidence clause	Quarantine output and reverify	Fabricated evidence

Legacy Operations Scenario

Separate source-discovery cohorts may run concurrently, while the shared synthesis register remains under orchestrator ownership.

Companion Skills

• research-orchestration supplies research wave structure.
• source-evaluation supplies mandatory evidence constraints.
• evidence-claim-graph supplies merge substrate.
• skill-writing applies when agent operations are encoded into skills.

Workflow

1. Partition independent work and reserve shared files for the orchestrator.
2. Brief each agent with exact ownership, inputs, outputs, permissions, and the evidence clause.
3. Stop a cohort when evidence rules are violated, ownership overlaps, or required inputs are absent.
4. Recover by quarantining suspect output, repairing the brief, and retrying only the failed bounded task.
5. Verify checkpoint evidence before merging results into the shared synthesis substrate.

Outputs

Artefact	Consumer	Acceptance condition
Agent briefs, checkpoints, and verified merge packet	Orchestrator and synthesis workflow	Ownership is non-overlapping and every accepted claim retains its evidence trail

Worked Example

Two source-discovery cohorts may run concurrently, but the orchestrator alone edits the shared claim register and rejects any checkpoint missing source links.