MANTA: Self-Evolving Organizational Topology for Multi-Agent Systems at Runtime

Core Claim

MANTA (Multi-Agent Network Topology Adaptation) argues that topology is not a static design-time choice, but an object that can self-evolve at execution time. When a multi-agent LLM system fails, most fixes target prompts, reasoning chains, or memories—but the organizational structure itself is the variable everyone ignores.

The Seven-Level Self-Improvement Pyramid

• L0 Outputs: Self-Refine — fix the answer
• L1 Prompts: DSPy, APE, TextGrad — fix instructions
• L2 Reasoning traces: CoT, ToT, GoT — fix chain-of-thought structure
• L3 Skills/tools: ReAct, Voyager — fix what can be called
• L4 Memory: Reflexion, ACE, A-MEM — fix what is remembered
• L5 Agent roles: CAMEL, MetaGPT, Debate — fix division of labor
• L6 Topology: MASS, AFlow, ADAS, MANTA — fix who talks to whom
• L7 Weights: RLHF — fix the model itself

Before MANTA, all L6 methods (MASS, AFlow, ADAS, AgentSquare) were offline optimizations—search for a good workflow, then freeze it. MANTA's single difference: topology can be modified during execution.

Architecture: Three Components, Two Decisions

A nested design: an inner multi-agent system solves the task; an outer orchestration layer shapes the inner structure using three LLM components:

1. Topology Planner — produces a compact topology plan from task features and experience memory; a deterministic expander unrolls it (roles, context policy, communication edges, visibility) and validates structural constraints. The planner never sees benchmark identity or reference topologies. 2. Trace Auditor — scans structured execution traces (tool records, relay packets, confidence, open issues, evidence visibility). It looks only at process, never answers, and cannot consult ground truth. If at least one fixable medium/high-severity flag appears, the run is "flagged" and triggers repair; otherwise it passes as "clean." 3. Controller — decides whether to finalize or apply a bounded revision: at most one revision, at most three operations (add agent, expand agent, regroup, change edges, change visibility).

The bounded design keeps meta-layer overhead at only 12% of the inference budget while still catching most repairable failures.

Two Playbooks

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

Critically, the long-term learning signal never sees ground truth. Each run receives only a process-derived label—"procedurally clean" (no auditor flags plus decision-level consensus)—which means no coordination anomalies were observed, not that the answer was correct. Ground truth is used only for evaluation.

Four Failures, Four Structural Fixes

1. Overloaded branch: a worker juggling multiple directions fails; the fix expands it into a dedicated subgroup, turning a star into a two-level tree. 2. Missing check (Math500): a lone low-confidence answer; the fix adds one critic agent paired into debate—without ever attempting to solve the problem itself. Next round: consensus confidence 1.0, score 1.0. 3. Premature consensus (PlanCraft): three agents agree too early with unresolved issues; the fix keeps all three but adds a direct worker–verifier edge (star → full debate). Next round passes. 4. Duplicated action: the same stateful action executed twice in parallel; the fix serializes execution.

All four fix structure, not content—who talks to whom, who checks whom, who goes first.

Results: 74.0, a 5.8-Point Lead

Five benchmarks (BrowseComp, StableToolBench, PlanCraft, WorkBench, MATH), three independent runs of 30 items each, Gemma backbone. MANTA averages 74.0 vs. 68.2 for the next best method.

| Configuration | Success rate | Total tokens | |---|---|---| | Full MANTA | 71.7% | 100,315 | | No initial Topology Planner | 57.5% | 112,040 | | No topology revision | 60.8% | 72,105 | | No long-term playbook update | 67.5% | 102,620 | | No long-term playbook | 66.7% | 78,356 |

Initial planning (-14.2 pts when removed) and runtime revision (-10.9 pts) contribute most; cross-run learning adds 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 shows catastrophic negative transfer—a workflow optimized for one task collapses on another. MANTA is the only method with positive average transfer because its playbook stores inheritable, actionable structural rules, not fixed workflows tuned to predefined tasks.

Key Insight: Procedurally Clean ≠ Correct

A run can be process-clean but answer-wrong, or process-messy but answer-right. Process-level auditing and answer-level correctness are orthogonal metrics; measuring only one misses failures in the other. MANTA chooses process-level signals because answer-level supervision is unavailable in real deployment—pragmatic, but it means MANTA learns "how to avoid process anomalies," not "how to produce correct answers."

Why Now

MANTA became feasible because (1) LLM calls are cheap enough that a 12% meta-layer budget is acceptable; (2) structured agent communication protocols (relay packets, append-only evidence ledgers, context policies) have matured; (3) machine-readable process traces make answer-free auditing possible. This explains why L6 was the last self-improvement level to be unlocked.

Limitations

• Only one revision is allowed—tasks needing multiple reorganizations are out of reach.
• Process-level signals have blind spots; failures outside the auditor's observable flag set go undetected.
• Single backbone model (Gemma); meta-layer quality depends on it.
• No open-source code released.

Takeaway

Organizational structure deserves first-class design status in the agent era: it is an optimizable variable whose granularity aligns with task granularity. After MANTA fills L6, the next frontier is making topology learning itself self-evolving.

---

Paper: arXiv 2607.28527 — MANTA: Multi-Agent Network Topology Adaptation for Self-Evolving Multi-Agent Systems

Code: No accompanying open-source repository.