1/ models are getting better at code! how you use agents can also have 10x impact. who is using coding agents internally in "advanced mode"? dm for @conviction-hosted dinner invite on this topic some notes from conversations w/ @conviction's startup teams about best practices:

2/ front-load context. keep an in the repo with project purpose, architectural guidance, file structure, style guide, and test commands so agents write to your standards

3/ phase the work. plan → implement → review → refine --> commit. have the agent explain design choices and catch bad assumptions early

4/ build/use validation in the loop and ongoing across your code base! from linters to , review is the missing link, and there's big upside in better automated validation

5/ create checkpoints -- you can outpace your own understanding of the codebase. the risk is everyone's "lgtm" and you have features no one can fully explain, debug, or change safely

6/ give your agents team memory. shared prompt templates + commands (/plan, /spec, /implement, /review, /revise, /commit) keep usage consistent and quality compounding

7/ pair with tooling like @CorridorSecure, and feed traces, scans and errors automatically back into the agent

8/ re-baseline automatically. when large changes happen, script a refresh of and a new codebase summary to keep the agent’s “mental model” current

9/9 parallelize intentionally. break work into scoped, independent tickets with tests + context. run agents in parallel and keep humans on design, integration, tricky 2nd-pass debugging, and critical reviews. don't watch the agents run!