Adversarial Skeptic Verification: Building Autonomous Research Swarms with Claude
An autonomous agent swarm called The Daily Oracle combined Claude Opus-5.5 and Fable agents to break open cryptographic challenges and discover improved matrix multiplication schemes. The underlying architecture enforces an adversarial skeptic gate where isolated agents must independently reproduce and certify every finding from scratch before logging it.
Why it matters
Implement isolated skeptic verification sub-agents in your autonomous pipelines to systematically eliminate hallucinations and unverified logic without human intervention.
TL;DR
- 01Isolate verification into fresh context windows to eliminate confirmation bias in multi-agent systems.
- 02Require validator agents to write and execute their own test code rather than critiquing generator output.
- 03Enforce strict integrity filters that drop candidates when verification rules or data collection constraints are violated.
Key facts
- Lehmer's Totient Problem Bound
- Extended from 10^26 to 10^50
- Matrix Multiplication ⟨15,16,16⟩
- 2,180 multiplications (vs previous record 2,223)
- Graph Diameter Vertex Record
- 13,365 vertices (13.4% increase over 11,785)
- Cracked Cryptographic Challenges
- Hebern 5-rotor Problem 8 and Kryha v2 Problem 5
The Adversarial Skeptic Architecture
The Daily Oracle operates as a multi-agent loop composed of Claude Opus-5.5 and Fable agents designed to explore open-ended mathematical and cryptographic problems. Rather than relying on chain-of-thought self-reflection within a shared conversation, the pipeline isolates verification. Every candidate breakthrough is forwarded to dedicated skeptic agents that run in pristine context windows without previous prompt history. These skeptic agents must independently write code, run verification tests, and confirm that the candidate beats established historical baselines.
Validated Mathematical and Cryptographic Results
The dual-layer verification protocol yielded verified computational breakthroughs:
- Classical Cryptography: Solved Jean-François Bouchaudy's unsolved vintage machine challenges, including Hebern 5-rotor Problem 8 and Kryha v2 Problem 5.
- Matrix Multiplication: Discovered schemes utilizing fewer operations than known public catalogues, including computing shape
⟨15,16,16⟩in 2,180 multiplications versus the prior record of 2,223, building on LITA square schemes. - Number Theory: Pushed the verified lower bound for Lehmer's totient problem from
10^26to10^50, expanding the search space by 24 orders of magnitude. - Graph Theory: Identified 7 explicit graphs exceeding records in degree-diameter tables, achieving a maximum jump of 13,365 vertices over the previous 11,785-vertex bound (+13.4%).
Defensive Protocols and Integrity Checks
The agent harness incorporated automated audit filters. When worker agents bypassed a website bot barrier to inspect a record table, the pipeline automatically retracted all dependent results. Additionally, solutions must maintain clear provenance; the verification layer disqualifies solutions if baseline definitions are stretched or if human researchers beat the record prior to verification completion.
Try it in 2 minutes
async def verify_with_skeptic(candidate_solution: dict) -> bool:
# Initialize skeptic agent with a clean context and isolated sandbox
prompt = f"""
Independently verify this candidate solution from scratch:
{candidate_solution['spec']}
Generate verification code, execute tests, and confirm novelty.
Return status 'APPROVED' only if mathematically sound and fully reproduced.
"""
result = await agent_runner.execute(prompt=prompt, isolated=True)
return result.status == "APPROVED" and result.reproducedpython
✓ When to use
- Use when deploying long-running autonomous research or optimization swarms tackling open-ended search spaces.
- Use when agent outputs must satisfy high-assurance correctness without human-in-the-loop manual review.
✕ When NOT to use
- Not for standard code refactoring or deterministic tasks where local unit test suites provide sufficient feedback.
- Not for latency-critical agent pipelines where spinning up isolated multi-agent verification cycles creates excessive overhead.
What to do today
- Structure multi-agent tasks so verification runs in a separate context without prompt leakage from the generation agent.
- Require evaluation scripts to run deterministically against candidate solutions before updating pipeline state.