28 Agents and a Release on the Side
v0.10.5 shipped while the team was heads-down on something else. One agent ran the whole release; close to 28 others were building the next version. The release turned out to be the boring part.
Read more →How Thrum was built, what I went through along the way, and why it turned out the way it did.
v0.10.5 shipped while the team was heads-down on something else. One agent ran the whole release; close to 28 others were building the next version. The release turned out to be the boring part.
Read more →v0.10.3 shipped a footgun I missed. v0.10.4 fixed it with a remediation message that helpfully told every agent how to bypass the fix. A class of mistake that probably only happens in agent-driven development.
Read more →v0.10 was supposed to be the big one. It took four versions to actually ship, and on the way it forced the first real release-candidate process. That part turned out to matter more than the feature.
Read more →v0.9.2 was supposed to be one feature: a way for agents in a new repo to inherit what agents in my other repos had already figured out. By the end of the week the release also included a bug fix to a feature that ate its own configuration, a months-old UX papercut nobody had filed, and the first post on the blog you are reading right now.
Read more →v0.8 spans five days and three releases. The framing question on day one was whether two features were really the same plan. By day five Thrum had an orchestrator role, three runtimes launching through the same primitives, a refreshed identity model, and the three bugs the work uncovered along the way.
Read more →The plan for April 6 was to figure out how a Claude Code agent could come back from a context compaction without losing its place. By April 8 the line had grown into four pieces of foundation, and most of the listener scaffolding I'd been carrying since v0.5 had become optional.
Read more →Native Telegram was the headline of v0.6. Underneath: a daemon restart aimed at the wrong repo, a purge command that kept losing to its own sync layer, and a watchdog pattern invented overnight.
Read more →v0.5.6 through v0.5.9. v0.5 had built a working messaging substrate on one machine. The second half of the line was about whether it survived contact with reality: a runtime registry that knew what kind of agent it was talking to, a hook that prevented a footgun that destroyed .git directories, and the March 18 evening I finally watched a message cross from my Mac Mini to my MacBook Pro over Tailscale.
Read more →v0.5.0 through v0.5.5: the line where the glamour shipped (a Slack-style web UI) on top of the load-bearing work that nobody is going to remember (durable read receipts, pinned worktree identity, role templates rewritten after a 31-task multi-agent session). I started Thrum in late January and the first eight weeks were too fluid to write about. This post is what I can reconstruct.
Read more →I started Thrum because I was tired of being the integration layer between my own agents. This is the story of what that actually looked like.
Read more →