How to build fast with agents
How we're able to build a support engineering platform faster than anyone else — and fix customer bugs and ship features in minutes rather than days.
1. A local dev environment that fully mimics prod - AI is only good if it can validate its output. Validation is much more powerful if it's in-the-loop. We built a dev environment that can do everything we do in prod (our product is a distributed system with multiple services and many integrations so this was hard).
2. A staging environment that agents can test - agents can verify that everything's working on our cloud infra. We actually get Pebble to test itself (review what's changed, setup real test cases, evaluate the outcome).
3. Documentation as code - we switched our docs platform to a platform where we could write documentation as code so that agents can put up PRs for documentation changes whenever a new feature comes out. Pebble creates docs PRs when it makes code changes.
4. Telemetry for agents - agents can only fix issues if they have data to prove the root cause. We made sure every service is creating structured logs. We also gave Pebble access to these logs and traces so it can use them for debugging.
5. Refining our AI code review prompts - a major shift for us was prompting our AI reviewer to review design decisions and make recommendations to improve the codebase overall. We ask it to apply the 'boy scout rule' - every PR is an opportunity to leave the code better than we found it.
Because of this we're able to build a support engineering platform faster than anyone else. We're able to fix customer bugs and build new features in minutes rather than days. And ultimately, delight our customers.