Topology selection
tnsai-coordination ships eight groups/ topologies. The trap is picking the one that matches an org chart. For a novel problem that needs continuous reasoning, that choice is usually wrong: each hop compresses context, and the whole drifts even when every node looks fine in isolation.
The reasoning behind this page is in the wiki article Context Engineering vs Role-Based. Topologies covers how each topology works. ROMA is a recursive solver, not an organization spec; see Recursive task solver.
Decision matrix
| Task | Use | Avoid |
|---|---|---|
| Sequential deterministic stages (ETL, CI, triage) | pipeline | network, matrix |
| Independent parallel hypotheses (research, search) | swarm, network, or RecursiveTaskSolver (ROMA) | pipeline, team |
| One coordinator with known workers | hubspoke | hierarchy unless a supervision tree adds real gates |
| Novel problem, reasoning must span the whole task | single agent + context engineering, or ROMA | team, hierarchy, pipeline |
| Dynamic membership + consensus | Council and voting | team |
| Quality gate | Judge | team with a reviewer role |
| Negotiation with incentives | Negotiation | debate used as a market |
| Adversarial / sanity check | Debate | council (council is consensus, not opposition) |
| Fixed steps with known dependencies | Workflows | a group topology |
| The same task sent to several agents | Fan-out | team |
coalition and matrix are for dynamic alliances and multi-dimension
assignment. They are not a default for “we have several specialists.”
When team, hierarchy, and pipeline fail
These three map onto the role-playing anti-pattern when the task is not a repeatable workflow with deterministic stage boundaries.
- Handoff compression. A 50-line chain of thought becomes a 3-line summary at the next node.
- Fake specialization.
LEAD/REVIEWERare prompt labels, not capacity splits. The same model sits behind each role. - Every node can pass while the whole drifts. Isolated tests stay green, but the integrated answer wanders.
They are the right tool for ETL, approval chains, and other pipelines whose stages have fixed contracts. They are the wrong tool for “figure this research question out.”
For novel work prefer:
- One
Agentwith tools, retrieval, and a verifier, with no group. com.tnsai.coordination.roma.RecursiveTaskSolverwhen you need parallel decomposition plus a verification gate that is more than a handoff.
When to choose ROMA
ROMA is the default multi-agent answer when:
- the task is not a pipeline with deterministic stages
- sub-problems can be explored independently
- you need a verifier that is not “the next role in the org chart”
Do not treat RecursiveTaskSolver as an org spec; the framework has no
organization type.