Anthropic Multi-Agent Workflow Uses 60 Subagents and Lean Verification for Mathematical Proofs
Anthropic deployed an experimental unreleased model across 60 specialized subagents to tackle the Riemann hypothesis. The multi-agent architecture spent 31 million tokens over 36 hours and verified generated results using the Lean proof assistant.
Impact: High
Why it matters
Demonstrates how decoupling task orchestration into exploration, validation, and writing agents paired with deterministic verification tools like Lean eliminates hallucinated logic in complex reasoning workflows.
TL;DR
- 01Decouple multi-agent workflows into distinct exploration and validation roles.
- 02Pair high-token LLM agent runs with programmatic code checkers like Lean to eliminate hallucinations.
- 03Expect heavy output token expenditure (30M+ tokens) for long-horizon autonomous tasks.
Key facts
- Total Subagents
- 60 agents
- Total Output Tokens
- 31M tokens
- Execution Runtime
- 36 hours (1.5 days)
- Ideas Tested
- 650 ideas
- Verification System
- Lean (open-source proof assistant)
Multi-Agent Subagent Architecture Breakdown
Anthropic details a 60-subagent hierarchy engineered to sustain reasoning over a 36-hour runtime:
2 idea generatorsresponsible for main core mathematical hypotheses.13 idea contributorsproviding supporting domain arguments.30 exploratory agentsattempting novel approaches (all failed, filtering bad paths).13 validator agentsperforming verification on step-by-step logic.2 paper writerscompiling formal output artifacts.
Deterministic Verification with Lean Proof Assistant
To prevent agent hallucinations from poisoning long context windows, Anthropic integrated the open-source Lean proof assistant into the loop. Agents translated proposed mathematical steps into Lean code, which provided objective pass/fail execution feedback.
Compute Scale and Token Consumption
Across 1.5 days of execution, the agent cluster consumed 31M output tokens while exploring 650 distinct candidate ideas. Human in-house mathematicians subsequently verified the Lean output, confirming valid lower-bound progress.
Try it in 2 minutes
# Example Lean 4 CLI setup for formal code / logic verification
elan self update && lake new proof_verifier math
cd proof_verifier && lake buildbash
✓ When to use
- Designing long-running multi-agent pipelines requiring zero tolerance for logic hallucinations
- Structuring subagent workflows with explicit generator and validator segregation
- Coupling LLM reasoning outputs with deterministic compilers, linters, or proof checkers
✕ When NOT to use
- Simple single-prompt tasks where token cost and execution latency must remain minimal
- Unstructured creative tasks lacking external deterministic programmatic evaluation
What to do today
- Integrate external CLI checkers or linters into agent loops to auto-reject invalid candidate outputs.
- Implement dedicated validator agent prompts separate from code generator agents in Claude Agent SDK workflows.
Sources