MANTA: Self-Evolving Multi-Agent Topologies at Runtime — What It Means

What MANTA Is

MANTA (Multi-Agent Network Topology Adaptation, arXiv 2607.28527) argues that topology is not a static design-time choice but a runtime self-evolving object. Where most LLM multi-agent work optimizes outputs (Self-Refine), prompts (DSPy), reasoning traces (CoT), tools (ReAct), memory (Reflexion), or roles (MetaGPT), prior topology methods (MASS, AFlow, ADAS, AgentSquare) all optimize offline — search a good workflow, then freeze it. MANTA instead modifies the topology during execution.

Architecture: Three Components, Two Decisions

A nested two-layer design: an inner target multi-agent system, and an outer orchestration layer with three LLM parts:

• Topology Planner — produces a compact topology plan from task features and experience memory; a deterministic expander unfolds it into roles, context policies, communication edges, and visibility policies. It never sees benchmark identity or hand-designed topologies.
• Trace Auditor — after each round, scans structured traces (tool logs, relay packets, confidence, open questions, evidence visibility). It inspects process only, never ground-truth answers. If at least one fixable medium/high-severity flag exists, the run is flagged for repair; otherwise it passes as clean.
• Controller — decides whether to finalize or apply one bounded topology revision: at most one revision, at most three operations (add agent, expand agent, change grouping, change edges, change visibility), then re-run and re-audit.

This boundedness keeps the meta-layer at only 12% of the inference budget.

Two Playbooks, No Ground Truth

• Short-term playbook (per-run): current topology, auditor flags, recommended fixes, controller decisions.
• Long-term playbook (cross-run): general rules of the form *task features + process risk → topology choice*, rewritten periodically by a Skill Reflector.

Critically, the long-term playbook's learning signal never consults ground truth. Runs receive only a process-derived label — 'procedurally clean' — meaning no coordination anomalies were observed, not that the answer was correct. This avoids training the process against answers.

Four Failures, Four Structural Fixes

1. Overloaded branch — one worker handling multiple aspects plus an unverified tool failure → expand the worker into a dedicated subgroup (star becomes a two-level tree). 2. Missing check (MATH500) — low-confidence single-agent answer → add a critic agent and pair them in debate. Next round: consensus confidence 1.0, score 1.0. 3. Premature consensus (PlanCraft) — three agents agreed round one with open questions → add a direct worker–verifier edge (star becomes full debate). Next round: 1.0 / 1.0. 4. Duplicated action — the same stateful action executed twice in parallel → convert parallel execution to a serial chain.

Every fix changes structure, not content — who talks to whom, who checks whom, what runs first.

Results

Five benchmarks, three independent runs of 30 items each, Gemma 4 backbone: MANTA averages 74.0 vs. 68.2 for the next-best method.

Ablations:

| Configuration | Success rate | Total tokens | |---|---|---| | Full MANTA | 71.7% | 100,315 | | Without initial Topology Planner | 57.5% | 112,040 | | Without topology revision | 60.8% | 72,105 | | Without long-term playbook updates | 67.5% | 102,620 | | Without long-term playbook | 66.7% | 78,356 |

Initial planning (-14.2 pts) and runtime revision (-10.9 pts) matter most; cross-run learning contributes 4–5 points but enables cross-domain transfer.

Cross-Domain Transfer: The Only Positive Case

| Method | PlanCraft→WorkBench | WorkBench→PlanCraft | Avg change | |---|---|---|---| | ADAS | 66.7→66.7 | 76.7→70.0 | -3.3 | | AgentSquare | 60.0→36.7 | 26.7→23.3 | -13.3 | | MASS | 46.7→0.0 | 70.0→0.0 | -58.3 | | MANTA | 43.3→50.0 | 70.0→70.0 | +3.3 |

MASS's fixed workflow learned on PlanCraft collapses completely on WorkBench. MANTA transfers because its playbook stores inheritable, actionable structural rules, not task-specific workflows.

Key Insight: 'Procedurally Clean ≠ Correct'

Process-level audit and answer-level correctness are orthogonal metrics. A run can be procedurally clean yet wrong (smooth coordination, insufficient knowledge), or messy yet right. MANTA uses process-level signals because ground-truth answers are unavailable in real deployment — a pragmatic choice that also means it learns "how to avoid process anomalies," not "how to be correct." That gap is future work.

Why Now

1. Cheap LLMs: the meta-layer fits in 12% of budget — unaffordable in the GPT-3 era. 2. Mature communication protocols: relay packets, append-only evidence ledgers, context policies. 3. Machine-readable process representations enabling answer-blind auditing.

This explains why topology (level 6 of the self-improvement pyramid: outputs → prompts → traces → tools → memory → roles → topology → weights) was the last level to be addressed.

Limitations

• Only one revision per run — tasks needing repeated reorganization are out of reach.
• Process-level signals have blind spots beyond the Auditor's flag set.
• Single backbone model (Gemma 4); meta-layer quality depends on it.
• No open-source code released (arXiv 2607.28527).

Takeaway

Organizational structure should be treated as a first-class, optimizable variable in agent engineering — a new dimension alongside context engineering (what each agent sees) and protocols (MCP/A2A). MANTA fills the previously empty 'L6 Topology' slot in the self-improvement hierarchy; the next open question is whether topology learning itself can self-evolve.

FAQ

Who is this for? Practitioners, researchers, and students interested in AI, machine learning, and LLM-based multi-agent systems.

Where is the code? The paper does not include a public repository. Paper reference: arXiv 2607.28527.