We’re back! 🎙️
Season 28 is all about building with AI: what happens when an agent deletes your database, why your app charged someone twice, and how to fix it.
Hosted by @lolocoding, sponsored by @TimescaleDB. Episode 1 drops Dec 2!
🎧 codenewbie.org
Here’s my coding agent glossary. I’m trying to clarify what these terms mean so that we can meaningfully discuss them.
Coding agent: the thing you directly prompt that has a single context window, system prompt, tool definitions, etc. Examples are Claude Code, OpenCode, Pi, Cursor Agent, Warp Agent.
Cloud coding agent: a coding agent that runs in a cloud sandbox, probably with some entrypoint through Slack or Jira. Examples are Devin, Cursor Cloud Agents.
Cloud coding agent automation: a trigger that invokes a cloud coding agent with a prompt (e.g. “When this Sentry event fires, try to fix the underlying crash.” Examples are Devin automations, Cursor automations.
Software factory: A code-defined system for a fully closed loop encoding all of your organization’s cloud coding agent workflows, plus all of the infra for measuring, improving and securing those workflows at scale. This is Warp Factories. It’s also what companies like Uber and Stripe are building internally and what I hear platform teams ask for every day.
It’s only a “software factory” imo if it’s headless infrastructure:
1/ All your agent workflows are defined in code so they can be versioned, tested and self-improved
2/ It runs in the cloud - otherwise you can’t govern it, or automate around it
3/ It supports running different coding agents (it’s a higher abstraction layer than “coding agent” – your factory should be able to run any agent to test how it performs)
4/ It stores all of the agent data for use in self-improvement
5/ It is measurable, testable and self-improving
A lot of folks are marketing “software factories” that are really just strings of prompts typing together local or cloud coding agents, which influential devs like @poteto rightfully dismiss this as marketing bunk.
I agree this isn’t an interesting approach, but I also think it’s not really what platform teams who care about increasing productivity while minimizing cost are trying to build.
I built the custom webhooks feature in Warp and set up the self-improvement loop with Sentry errors -> fixes within Warp 😄
Check out this article to see what else we're doing for self-improvement!
I've been thinking about what a true, closed-loop software factory will look like.
One where agents know what to fix, what costs to optimize, and the big one: what features to actually build.
We're not there yet. But maybe we're getting close...
What's the fastest way to get a hold of good support at Meta these days?
A small business I care about deeply is dealing with their Facebook page being taken over and is trying to figure out ways to remediate.
Got to spend some time back in the Azure Portal while building this out...very nostalgic!
Check it out -- I'm particularly happy with the automations support we landed here.
Warp is one of the few (only?) cloud agent platforms that lets you integrate with ADO's events so easily.
Warp Factories now supports Microsoft Teams and Azure DevOps.
Software factories should be built on flexible infrastructure to run agents from anywhere. Thats why we’re the first in the industry to support the full Microsoft suite across software factory workflows.
Moving to a model where engineers can "self-stamp" PRs authored by agents they initiated is an easy way to reduce the code review burden on a team.
It also progresses nicely to the model where agents are fired autonomously to make code changes in reaction to bug reports/crashes/etc. instead of human intent.
I sat down with @clairevo to talk about how we're automating development at Warp.
One big change is how we approach code review - for low risk changes, we now let engineers review agent PRs and ship themselves, rather than requiring a second dev’s review.
We are investing more and more into tooling to understand and optimize our coding agents like benchmarks, agent tracing, etc to figure out why are they spending so much, how to increase prompt adherence etc. It's the next gen of "devtools."
it's interesting because we are also investing at the same time in ways to get agents to use the old generation of devtools (e.g. profilers, debuggers, chrome devtools) that humans used to use to debug code
we just keep going up in abstraction layers but it's the same stuff
Introducing Factory Benchmarks: The first model bench generated from your own coding tasks.
Measure, test and improve coding agents by replaying past agent runs, and cut cost-per-PR by 63%+
Here’s how it works 🧵
I'll be speaking at @kc_dc next week -- reprising my "building a web framework from scratch" talk for the last time.
It's 1-hour of AI-free, slide-free, live-coding where I start with a HTTP server and build out the core components of an API framework in C#.
End of an era 🥲
We see everyday how much the @warpdotdev team cares.
But what was special to me here was being able to help @warpdotdev build faster, and then help Warp's customers build faster too.
Warp started by building the terminal every developer loves. Today, their products reach 51% of the Fortune 500.
Now Warp uses Namespace to build more, faster. With Namespace, they've 4x'd their CI performance and now run 20,000+ sandboxes a month for their cloud agent platform.
We went to NYC to sit down with Aloke, Head of Product Engineering at Warp, to get the full story.