saved
The road to self-improving software
Hraness wrote this summary from a saved copy of the source. Quotations are taken word for word from the source.
gist
In an X Article, Ben Holmes argues that a true software factory closes one large loop from task to shipped product, with agents deciding what to fix and build next rather than just producing pull requests faster. He walks through three loops Warp is deploying: agents fixing alerted bugs, agents scoring their own sessions to improve skills, and agents mining customer interviews for roadmap items.
ideas
- Fill the backlog, not just work it. Error alerts from tools like Sentry trigger an always-on agent through a webhook, which investigates and opens a draft fix PR for human review.
- Factory engineering. Scorer agents grade past sessions on rubrics such as efficiency, and a self-improvement agent turns common failures into skill or AGENTS.md changes.
- Improve the process, not only the code. Cheaper, faster changes indirectly raise velocity and quality and build enough trust to merge with less line-by-line scrutiny.
- Agents on the roadmap. Customer-call transcripts from Grain, exposed over MCP, are mined for recurring friction and turned into Slack digests and Notion tasks that humans still approve.
quotes
“I've been thinking about what a true, closed-loop software factory will look like.”
“How do we get agents to proactively decide what to build next?”
“The agent should fix its own bugs.”
“I agree a human will always be in the loop for product work.”