Lartey Doryumu @LDoryumu - Twitter Profile

LDoryumu retweeted

5 days ago

We don't think AI agents should get full repo access from day one. They should earn it, much like a new engineer joining the team. We're building an AI Factory, with Velocity Core handling multi-agent orchestration and execution. The idea is to automate more of the software development process, from engineering tickets to pull requests, without losing control over what actually ships. One question we've spent time thinking about is how much autonomy these agents should have. Here's the model we're designing: 1. Start with a plan. The agent reads the ticket, writes an implementation plan and estimates the work. No code changes yet. This is also where we start if a repository doesn't have reliable builds and tests. 2. Open a draft PR. This is the default. The agent writes code and opens a draft pull request. An engineer reviews the work and decides when it's ready. 3. Open a ready PR. Once the agent has proven itself, it can skip the draft. Our proposed threshold is 70% of runs passing CI on the first attempt, at least 50% of PRs merging without changes, and zero reverts in the last 20 runs. 4. Assign a reviewer automatically. If that performance holds over an agreed period, the agent can open a ready PR and assign a reviewer directly. A person still approves and merges the code. 5. Eventually, auto-merge. Not in the first version. Further down the road, we may allow it for specific, low-risk tasks, and only with explicit client approval. The important part is how agents move between these levels. Moving up requires human approval. Moving down happens automatically. One revert, or repeated attempts to make a forbidden change, and the agent loses a level of autonomy. And trust is earned separately for each repository. Being reliable in one codebase doesn't automatically make an agent trustworthy in another. That's the approach we're taking with our AI Factory: give agents more responsibility as they demonstrate they can handle it, rather than giving them permissions just because we can. Where would you draw the line? Would you let an AI agent open a ready PR without a human reviewing the draft first?

43

72

16

47

12K