← Back to Field Notes

Series · 5 notes

Hosted agent runs

The agents we host: the engine, the runner, its network, its models, and what broke on the way.

  1. Part 1 Log Aug 10 – 16, 2026

    Orca-strator's first two days: a workflow builder that measured before it believed

    On August 14 we started Orca-strator, a local-first visual builder for AI workflows. By the end of the next day it had 15 merged pull requests and a handoff document whose first rule is that every load-bearing claim was wrong until it was measured.

  2. Part 2 Deep dive Aug 14 – 28, 2026

    Unknown is never zero

    In Orca-strator's first two days, token counts, prices and a planner's spend all read zero when the truth was unknown, and a budget built on them could never fire. Two weeks later our first hosted agents reported success after reading nothing at all.

  3. Part 3 Deep dive Aug 24 – 30, 2026

    A kill switch the customer holds: how infinary_controls pairs with our hub

    To let hosted AI agents into a customer's ERP we wrote infinary_controls, a Frappe app on the customer's side. Access starts from a code the customer pastes in, every request must be signed by us, and the off switch works without asking us.

  4. Part 4 Deep dive Oct 15 – 20, 2026

    Hosted agent runs move from Gemini to Claude

    Our agent catalog had named Claude models since October 15, but the hosted runner only held a Gemini key, so the workers most likely ran on Gemini. On October 20 the runner got an Anthropic key, and the first Claude runs turned up two gaps we fixed that night.

  5. Part 5 Deep dive Oct 19 – 20, 2026

    The runner's egress wait, and the clock we anchored it to

    About one hosted agent run in three couldn't reach the customer's ERP. Every failure came in the run's first minute. Our first fix measured its 120-second wait from process start, a clock that over-reports on Cloud Run. The second fix starts the clock when the wait begins.