Six months of writing code exclusively with agents

Who Should Read

A backend developer running coding agents in parallel who is hitting validation and review bottlenecks

Core Mechanics

• He banned hand-written code
  • In February 2026 he stopped writing code himself and fixed what the agent lacked instead of finishing for it.
  • He broke the rule once for three minutes and quit after seeing how much typing remained.
• Parallel agents collided
  • On one machine they fought over files, Git state, ports and processes.
  • Worktrees fixed only Git, AGENTS.md rules ate context, containers leaked, so each task got its own VM.
• YOLO mode inside disposable VMs
  • Approving every tool call made him the bottleneck, so agents ran freely in throwaway VMs.
  • External access was read-only, and writes were limited to test environments.
  • A proxy injected credentials so the agent never saw secrets.
• Passing checks did not mean worth shipping
  • He threw away changes that had green tests, screenshots and review-agent approval.
  • Tools can say a change works but not whether it is worth adding.

Community Reactions

• A commenter hit a token limit, could not continue and felt skill atrophy
  • They judged the approach wrong after a couple of months and worried about outages or price increases.
• A commenter cuts review load with a different-family model plus deterministic gates
  • Same-family reviewers share bias, so a different model reviews and lint and unit tests run first.
  • Gates still cannot tell whether something is worth building.
• A commenter said parallel agents left them exhausted
  • If all agents choose slow paths it is a net loss.
  • Their personal estimate: about 5x for triage and debugging, 2-3x for coding with solid tests, -1 to 1.5x for greenfield without tests.
• A commenter using Cursor with Claude said specs take long and only 1 in 5 attempts works
  • Others ignore 80% of the spec or diverge badly.
• Many commenters thought the post read like an ad and AI-written
  • The exe.dev job mention and AI-style prose drew complaints.

How to Apply

• Hand a customer report verbatim to an agent
  • Today, paste one bug report unsummarized and have the agent investigate logs.
  • Give read-only log access and decide yourself from its evidence.
• Isolate each agent task in a disposable environment
  • Give each parallel agent its own VM or container.
  • Start with read-only external access and limit writes to test environments.
  • Have a proxy outside the agent inject API keys.
• Require design discussion and behavior tests before merge
  • Discuss architecture, interfaces and constraints with the agent before code.
  • Use contract or property tests that pin invariants instead of tests mirroring the implementation.
• Store agent conversation history in SQLite or similar
  • Save every session so another agent can later search where work got stuck.
  • Keep records in a store independent of any one tool.

Try It

# Paste into Claude or Claude Code (simplified version of the article's investigation prompt)

Customer reports: {{customer report, verbatim}}

Please figure out what happened using the {{log system, e.g. ClickHouse}} logs and the relevant code.

# Instructions
1. Do not summarize; investigate based on the customer's own wording.
2. List the evidence you found in logs and code.
3. Give the cause, whether it is working as intended, and fix options. I make the decision.

# Next to try
Ask it to first judge whether this is solved by a doc or an email instead of code.

Terminology

Related Papers

• Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy

  Saving long tool output to files cut GPT-5.1 accuracy by 11.3 points; naming the authoritative file restored it.

• Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents

  Fake headlines hijacked LLM trading agents 62% of the time versus 14% for prompt injection.

• Harness Engineering 101: The LLM is the brain, the harness is the body

  An agent is a loop managing a JSON array; 16 chapters build one from scratch

• Harness2 by Google Research – Recursive Agent Harnessing

  Recursively refining the harness with fixed model weights raised benchmark scores by up to 11.8 points.

• Ousted Savannah police officer gave ICE agent husband access to Flock system

  A police officer shared her Flock login with her ICE-agent husband; an audit found 983 searches and she was fired

• Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling

  Running Claude Code agents in tmux works well, but limits appear around four agents.

Related Resources

• https://blog.exe.dev/engineering-with-ai