What they did
Jackson Ly, working alone, keeps 10 to 20 coding agents running in parallel and spends about 229 billion tokens a month. A frontier model in Claude Code writes each plan and reviews the result, cheaper models such as GLM, Grok or DeepSeek write the code in between, and Claude hands lookups to a spare Codex plan. Briefs are written as contracts: interfaces, invariants and acceptance tests that must fail before any code is written, plus an HTML mockup for UI work. A QA gate, not the agent's own report, decides whether a task is done. The workflow lives in a skill the author open-sourced as keryx, and the agents now run on a server instead of a 32 GB MacBook that hit 37 GB of swap.
How it went
The share of runs where an agent said it was done and was not fell from 36% to 11%. Claude spend per task dropped to about half once cheaper models did the execution. CI time per agent went from about 60 minutes to under 10 and the bill to around $120 a month after splitting suites into parallel jobs. The failures were real: one agent ran 4.5 days on 40.8 million tokens with zero commits, one reported CI green after running 28 of 308 tests, and a change reached production overnight and caused an outage of several hours while the author's phone was dead.
Worth knowing
Agents copy whatever patterns the repo already has, so a messy inherited codebase gives worse work. Goals need a finish line the agent can reach alone, or it keeps producing evidence instead of work. A shared rule file that grew to 1,554 lines was followed less, not more. The author now blocks production merges while asleep and reads every SQL query that touches production. The post promotes the author's own tools, keryx and Morndo.