Programmer. Currently discovering how much I don't know about graph query engines, database internals, and all that jazz.
London
Joined September 2010
- Tweets1.5K
- Following451
- Followers163
- Likes398
1) I run paired agent sessions on the same working directory: an overseer that I talk to at a high altitude, and a developer that does the hands-on work.
I stay at the overseer's altitude, have side conversations mid-build without polluting the dev's context, tee up work, etc.
8) I run my agents across 3 separate machines so that concurrent heavy (rust) build + test runs don't hurt each other too much. The fleet HUD shows the state of all agents across my machines, and all builds + PR's in CI.
Someone in the Rhondda who bought a lotto ticket in June, you have until 3rd December to claim your £12m lotto squids!
Really hope whoever it is manages to find their ticket and claim the win!🤞 If its one of my old mates, you owe me a beer or several 🍻
why, oh why, does claude want so badly to create follow-up tickets instead of finishing the damn job?
My golden rules for agentic coding:
1. Don't be lazy
2. Be adversarial, question everything
3. Probe the negative space
4. Don't blame the tool
5. Own the big picture / raise your level of abstraction
6. Context is king
7. Code doesn't really matter*
* More of a guideline(?)
i've been really enjoying this agentic coding working pattern over the last 4-5 weeks, running two sessions at different altitude, where i mostly converse with the "overseer" at high level. i'm running claude opus in both, but could easily be different models.
it works great for being able to have side conversations while development work is in flight and without polluting the developer context or having to /rewind or whatever.
also use the overseer to create good /compact directives to manage the developer's context.
