Raltic field guide
Seven multi-agent workflow patterns that teams can review
Multi-agent systems are useful when roles are genuinely distinct. These seven patterns keep task ownership, context transfer, evidence, and review visible instead of simulating an organization chart.
Move from the definition to the operating decision without losing the thread.
Direct answer
The most useful multi-agent patterns are router-specialist, parallel research, researcher-critic, planner-executor, local-cloud split, exception escalation, and synthesis-review. Each pattern needs explicit task ownership, minimal context transfer, observable outputs, and a human owner for the consequential result.
1-2. Router-specialist and planner-executor
Use routing when tasks need different expertise; use planning when execution benefits from an explicit, reviewable task breakdown.
A router classifies the request and assigns it to one specialist. A planner decomposes a larger objective into tasks that executors claim and update. Keep the routing criteria and plan visible so humans can correct a bad decomposition early.
Do not let the router become an opaque gate. Record why a task was assigned and make it easy to reassign.
- 01Router-specialist: one clear owner per classified task
- 02Planner-executor: review the plan before expensive execution
- 03Use task states for handoffs
- 04Escalate ambiguous routing to a person
3-4. Parallel research and researcher-critic
Parallelize independent evidence domains, then use a critic to challenge synthesis rather than repeat the same search.
Parallel research works when agents receive non-overlapping scopes and a common citation standard. A researcher-critic pattern adds independent pressure testing: the critic looks for unsupported claims, missing counterevidence, and stale sources.
The synthesis should preserve disagreements instead of forcing false consensus.
- 01Partition evidence by domain or source type
- 02Use shared freshness and citation rules
- 03Give the critic access to the brief and evidence
- 04Record unresolved contradictions
5-6. Local-cloud split and exception escalation
Assign machine-bound work locally, managed research in the cloud, and uncertain exceptions to a human.
A cloud agent can gather public or connected evidence while a local runtime inspects a repository. Transfer only the findings needed for synthesis; do not move private context between runtimes by default.
Exception escalation is a powerful pattern: the agent handles routine cases but creates a review task when confidence is low, evidence conflicts, or a prohibited action would be required.
- 01Label each runtime and accessible context
- 02Transfer minimum necessary evidence
- 03Define confidence or rule-based escalation
- 04Keep external actions human-owned
7. Synthesis-review
A synthesizer combines specialist outputs, then a human reviews the decision packet against the original brief.
The synthesizer should not erase source attribution. It should identify agreements, contradictions, missing work, and the recommendation being proposed.
The human reviewer evaluates whether the final packet answers the original question and whether any additional evidence or authorized action is needed.
- 01Preserve specialist attribution
- 02Separate facts, inference, and recommendation
- 03Surface incomplete tasks
- 04Request one explicit decision
- 05Record approval or revision
Primary sources
Product and practice claims in this guide are grounded in current first-party or standards sources. Raltic product claims are limited to shipped behavior described on this site.
FAQ