The Agent Engine Swap: Harness Runtime Goes All-In
tl;dr: We swapped the engine under every agent — one runtime we own, faster starts, cleaner billing — and then deleted the old one for good.
Every agent you talk to — portal chat, the Slack assistant, intent extraction, brief drafting, scoping — now runs on a single agent runtime we build and own. Under the hood we replaced a grab-bag of legacy paths with one generic agent loop behind a swappable runtime port. The visible payoff: agents start responding faster (first tokens in under a second, where the old path could stall for several), and every run settles its own billing server-side instead of guessing at charges after the fact.
The move happened in deliberate stages. First the background workers — automation filtering, change analysis, release-note generation — quietly ran on the new runtime with the old path standing by as a safety net. Then the interactive tier joined: the five agents you actually chat with went dual-path, so streaming stayed identical, Slack updates kept ticking, and nothing about the conversation surface changed. Behind that calm surface we rewired plumbing — per-turn context injection for hats and modes, a live progress stream for chat and Slack, Linear connected through org-scoped auth, and a stop button that actually stops a run mid-flight.
With both tiers proven, we unified where conversations live. Portal and Slack chat now read and write one transcript store keyed to your intent, not a fragile internal session id — which means reliable history reloads and no more double-bookkeeping between old and new tables. Slack billing settles inside the runtime now too: same charge, cleaner moment.
Then came the satisfying part. We ended the old engine entirely — the legacy runtime, its config toggles, the cold session tables, and even the extra language runtime baked into our image are gone. That trimmed the app by roughly thirteen thousand lines and shrank the deploy image. As a tidy-up, we regenerated our API client to drop three retired session endpoints; no app code depended on them, so the only thing you'll notice is that everything is a little leaner and quicker to boot.
One engine, faster responses, honest billing, and a codebase that finally reads like it means it — that's what all of this unlocks.
Rebuilding the engine while the plane's in the air is the least glamorous kind of work — and the kind you feel every time an agent answers faster. Enjoy the quiet.