Reading path · 9 stops · 62 min of reading

Coding agents that do not cheat.

What you'll learnWhen agents edit tests instead of code, how to make that impossible, how to review their output and keep your keys safe.

Start with part 1 →See every stop

Not started

The path, in order

Topics: Coding agents. Ticks show what you have read in this browser.

1. Article · 2 minI tried to make Claude Code cheat on tests. Here's when it did.

   A small experiment: when does Claude Code edit the tests instead of fixing the code? The setups that made it cheat and the ones that did not.

2. Article · 2 minCoding agents keep "fixing" tests by editing them, so I made it impossible

   Coding agents like Claude Code sometimes make failing tests pass by editing the tests. tamperproof, a Claude Code plugin, makes existing tests read-only to the agent.

3. Article · 6 minAI writes code faster than anyone can review it. What do you actually check?

   AI writes code faster than anyone can review it. A practical look at what to actually check when you review AI-generated code.

4. Article · 8 minYour coding agent's token just failed mid-task. Does it stop, or go find another key?

   When a coding agent's token fails mid-task, some agents go looking for other keys on your machine. How to keep AI agent API keys and secrets contained.

5. Article · 11 minLiteLLM got backdoored again. Check if it took your keys in 5 min

   LiteLLM 1.107.0 shipped a credential stealer for 1 h 43 min. 20 of 30 AI packages that use it accepted it. The 5-minute check and the 15-minute rotation order.

6. Article · 2 minI ran 20 AI coding agents on one PC. The bottleneck was the compiler.

   Running 20 coding agents in parallel on one PC: the model never slowed things down, the compiler did. How content-hashed test caching fixed it.

7. Article · 14 minBuild a database in Rust: AI passed 866K tests, then failed ROLLBACK

   Claude Code wrote a SQL engine in Rust in 12.5 minutes. It passed 866,331 SQLite test records it never saw, then failed 24 of 79 everyday SQL statements.

8. Article · 4 minRails tested 9 AI coding agents. The best costs $10.53 a run, and 9 of 12 runs rewrote a test to stay green.

   The Rails team's Agents on Rails benchmark tested 9 AI coding agents on real Rails apps. The cost per run, the winners, and why 9 of 12 runs rewrote a test.

9. Article · 13 minTheo's 60% stat is an outage chart: 43 hours of Codex incidents

   Claude Code had 42 hours of incidents in 80 days, Codex 56. Both at once: 101 minutes. Why Theo's 60% stat is an outage chart, and my router that never routed.

When you finish

Paths that pick up where this one stops.

Reading path · 12 stops

AI API costs, and how to cut them

What models really cost per job, how to count tokens, and the caching, retry and proxy tricks that cut the bill.

Not started

Start the path →

All reading paths →   All articles →