Orchestration
Hours-or-days runs are not the market baseline
Everyone wants agents that run for days. The differentiator isn't a longer prompt — it's lifecycle orchestration that survives delivery state changing underneath the run.
June 11, 2026 · The AgentLoom team
Work that runs for hours or days is the promise everyone wants from coding agents — and it is not what the market baseline delivers. The baseline is the interactive session: useful, real, and bounded by three hard limits. Context windows fill. Sessions time out. And a human has to be present, because the session stops the moment it needs anything.
Why long runs actually fail
The naive fix is a longer session — more context, more patience, a bigger prompt. It fails for a structural reason: delivery state changes underneath the run. CI fails after the code was written. A reviewer blocks after CI went green. The merge conflicts after the review passed. Main goes red after the merge landed. A single session would have to hold all of that in one context while it decays — and context rot is real: the longer a session runs, the worse its judgment gets.
So the differentiator isn't model stamina. It's lifecycle orchestration: the ability to decompose delivery into many short, sharp sessions and carry the right state between them while the world changes.
What carrying state actually takes
Bounded sessions — each phase gets a fresh, focused context instead of one long degrading one.
Checkpoints — the run's position in the lifecycle survives CI waits, interruptions, and restarts; nothing is recomputed.
Runtime memory, scoped per issue — what the team tried, what CI said, what the reviewer demanded — carried forward across every resume, never bleeding in from unrelated work.
Structured artifacts — each session ends in an inspectable artifact the next session picks up, not a chat transcript.
A delivery state machine — resolve, CI fails, resolve again, review blocks, resolve again, merge conflict, post-merge red, main green. The loop keeps moving because every transition is a state, not a surprise.
There's a quiet economic property here too: a run that's checkpointed can sleep. While your CI spins for twenty minutes or a reviewer takes the afternoon, the run isn't burning tokens pretending to wait — it resumes when the result lands. Days-long runs are only honest if the idle hours cost nothing.
The proof we live in
AgentLoom is built through the same orchestrated workflow it sells — issues selected, teams composed, sessions run, PRs reviewed, main repaired, follow-ups filed. At the time of writing, that loop has shipped hundreds of resolved issues across hundreds of thousands of lines of code, running day after day. The numbers on our homepage come from the repo's own telemetry, refreshed before publication — operational proof, not vanity metrics.