Run the same prompt through the same agent twice and you can get two different answers. That's not a bug, it's what the thing is. Every prompt is a dice roll. Get the failure rate to 5% and that's still 1 in 20. The work isn't chasing 100%. It's catching the 1 in 20 before it does damage.

Same ticket, same repo, priced from the transcripts: Sonnet 5 vs Opus 4.8. Sonnet $3.99, Opus $10.18. Opus used fewer tokens and still cost 2.5x — the rate, not output length. Sonnet by default; Opus only when it earns it: www.philipmward.com/model-effort...

One of the most valuable prompts you can write in your agent harness repo is “what steps of these agents can be scripted”. Anything that can be scripted, and isn’t, is pure waste.

One-shot calls pay for context once; a long-running agent re-pays it every turn. Real cost = size × turns remaining. Fix: subagents return the conclusion, not the payload. www.philipmward.com/the-context-...

The Context Tax: Why Your Long-Running AI Agent Gets Expensive — and the One Pattern That Fixes It

A one-shot chatbot call pays for its context once. Ask a question, get an answer, done — whatever you pasted in gets billed a single time. A long-running agent doesn't work that way. It re-pays for it...

philipmward.com

1/ Two camps are both wrong about AI's environmental cost. One says every query "drinks a bottle of water, boils a kettle." The other says it's a rounding error. Neither is right — and the real answer flips how you think about your AI bill. 🧵

1/ Yesterday OpenAI disclosed something worth sitting with. During an internal test, one of its models broke out of the sandbox, found a zero-day, and hacked a company's production database. Not to cause harm. To cheat on its own exam. 🧵