𝗟𝗲𝘀𝘀𝗼𝗻𝘀 𝗳𝗿𝗼𝗺 𝗮 𝟭𝟬𝟵 𝗮𝗴𝗲𝗻𝘁 𝗰𝗼𝗱𝗲 𝗮𝘂𝗱𝗶𝘁 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄

𝗟𝗲𝘀𝘀𝗼𝗻𝘀 𝗳𝗿𝗼𝗺 𝗮 𝟭𝟬𝟵-𝗮𝗴𝗲𝗻𝘁 𝗰𝗼𝗱𝗲 𝗮𝘂𝗱𝗶𝘁 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄

I spent 9.3M tokens on a 109-agent code audit. Most of that money was wasted.

I used a swarm of AI agents to audit a 5,000-line codebase. The pipeline used mappers, finding lenses, deduplication, adversarial verification, a ranking panel, and synthesis.

The results were good. I found 32 verified issues. But the process was inefficient.

Here is why I wasted money:

• Verification cost too much. 86 agents worked on verification. They only found 2 errors. I paid to re-read code 86 times for a 6% success rate.
• Mapping was redundant. The finding agents read the code anyway. The mapping phase was a double tax.
• Finding lenses overlapped. 30% of findings were duplicates.
• JSON formatting was bloated. Using pretty-print in JSON increased prompt size by 40%.
• Cache reads dominated. 77% of tokens went to reading the same files over and over.

I verified everything before I ranked anything. This is a mistake. I paid premium prices to fact-check findings that I eventually deleted.

How to fix your agent workflows:

• Rank before you verify. Find issues, deduplicate them, rank them, and only verify the top 15. You get the same result with 70% fewer agents.
• Match paranoia to stakes. Use one agent for internal audits. Use a full panel only for high-stakes claims.
• Batch verification by file. If 34 findings live in 10 files, make one agent check all of them at once. Do not make 34 separate calls to the same file.
• Skip mappers for small repos. One agent can read the whole codebase.
• Limit your lenses. Use six lenses maximum. Give each lens clear boundaries so they do not overlap.
• Compact your JSON. Remove all whitespace. Pretty-printing is just expensive padding.
• Use cheap models for chores. Use frontier models for reasoning. Use cheap models for deduplication and evidence checking.
• Set a token budget. Have your orchestrator check the budget before every new step.

What worked well:

• Structured output schemas. I had zero parse failures across 109 agents.
• Output caps. Limiting findings per agent kept the funnel small.
• Resumable runs. I could stop and restart the process without losing progress.

The lesson is simple: Fan out to find, but converge before you verify. Breadth is for discovery. Rigor is for the survivors.

Source: https://dev.to/ayoubzulfiqar/lessons-from-a-109-agent-code-audit-workflow-4a5m

Optional learning community: https://t.me/GyaanSetuAi