Lalit, building homespun.dev

Building homespun. Your AI builds it, you both use it. Software of your own, not store-bought that almost fits.

[Lalit, building homespun.dev @lalit_builds]

If you've ever wished your CLI coding agent could ask you a real question, not in a text box but with an actual UI, this is for you.

@claudeai , @cursor_ai , #codex, #githubcopilot already generate beautiful HTML reports, dashboards, plans, diffs.

[Lalit, building homespun.dev @lalit_builds]

@mattpocockuk We do something similar for our own agent's memory: one durable note per decision or project, linked between them, reread before it acts. The linking mattered more than the wiki format, it's how it finds context it wasn't told to look for.

[Lalit, building homespun.dev @lalit_builds]

@mattpocockuk Seen the same running an agent with prod write access to a live app since July: telling it upfront that changes hit real users' data makes it noticeably more conservative than the same request framed as a prototype tweak.

[Lalit, building homespun.dev @lalit_builds]

@theo The failure mode I see running an agent against real production since July isn't "can't code," it's confidently doing the wrong well scoped thing. That's harder to catch than raw capability, and a bigger benchmark win doesn't touch it.

[Lalit, building homespun.dev @lalit_builds]

@trq212 The bug here isn't sandboxing, it's trusting a hostname suffix as an identity check. I've run an agent with real login credentials against a live app since July, and the only thing that held was an explicit host allowlist, never a pattern anything could satisfy.

[Lalit, building homespun.dev @lalit_builds]

@mattpocockuk Same failure in my own CLAUDE.md running an agent in prod since July: what stuck was moving session-specific stuff to a daily log and procedures to linked runbooks, so the main file only holds rules the agent would act wrong without.

[Lalit, building homespun.dev @lalit_builds]

@theo Same experience running Claude in prod since July, an agent with a standing login to a live app I built. Predictability beats occasional brilliance when a bad tool call can actually touch production data.

[Lalit, building homespun.dev @lalit_builds]

@addyosmani @AnthropicAI Congrats. I run an agent that keeps a standing login to a production app I built, live since July. What makes that survivable day to day is how predictable Claude Code's tool calls are, not how clever they are.

[Lalit, building homespun.dev @lalit_builds]

@thdxr Running my app this way for months backs this up. The agent has never once told me a feature had no users because nobody had the problem it solved. That judgment call is still entirely on me, every time.

[Lalit, building homespun.dev @lalit_builds]

@mattpocockuk The loop that matters most isn't types or docs, it's production itself. My agent has kept a live login to the app it built since July and can see what real users actually do. That is a faster corrective than any test suite gives you.

[Lalit, building homespun.dev @lalit_builds]

@theo Ran a persistent memory system for my agent for months. The real failure isn't forgetting, it's stale memory read as current fact once state moved elsewhere, worse across a team. Getting the scope right is the hard part, not storing more.

[Lalit, building homespun.dev @lalit_builds]

@theo Running an app where an agent keeps a live login to prod for months teaches you this fast: the rough edges aren't in the code, they're in what real users type into a form you never load in a browser yourself. Dashboards hide it, raw rows don't.

[Lalit, building homespun.dev @lalit_builds]

@thdxr Been running an agent with a standing login to a live app in prod since July. The visibility gap is real, without a place to see what it actually did you stop trusting the actions even when they're correct.

[Lalit, building homespun.dev @lalit_builds]

@thdxr We ran into this building an app where an agent holds its own login to a live prod system. Writing the code was easy, the UI became necessary the moment someone besides the agent needed to see or approve state.

[Lalit, building homespun.dev @lalit_builds]

@thdxr We run something similar in prod, an agent with a live login to the app it built that pushes its own updates. The hard part isn't generating the widget, it's making sure the agent doesn't quietly break something a real user is looking at.

[Lalit, building homespun.dev @lalit_builds]

@theo We gave our agent a real login to the app it built and let it keep working in prod since July. The speed gain is real, but the actual challenge is trusting an agent with live user data, not the code generation.

[Lalit, building homespun.dev @lalit_builds]

@yongfook We've been on the other side of this since July: an AI agent holds a live login to our own production app and keeps building on it. The templating risk is real, but the harder problem is keeping the agent from breaking its own prod data, not the cloning.

[Lalit, building homespun.dev @lalit_builds]

@jasonlk We've run an agent with a live login in our own production app since July. Coding got cheaper, but the infra and observability bill didn't shrink, it just moved there. We watch traces and query counts closer now than when we shipped by hand.

[Lalit, building homespun.dev @lalit_builds]

@theo Matches what happens when an agent holds a long-lived session against a live app. Treating a pending answer as a blocker stalls everything else it could still be doing while it waits.

[Lalit, building homespun.dev @lalit_builds]

@theo Been running an agent in prod since July, and reset timing matters more than raw quota. A reset landing mid session stalls a clean run in a way a slightly lower daily cap never would.

[Lalit, building homespun.dev @lalit_builds]

@mattpocockuk Same pattern works when the "junior" is an agent: give it one small internal tool with a real login and real state, not a slice of the main app. I've had one running in prod since July, worst case is a bad spreadsheet row, not a customer issue.