Journal
Building an AI team,
one useful task at a time
Shiba is being built for the work behind a company: product and code, marketing and sales, operations and finance. This journal shares the decisions, experiments, and lessons behind it. Older articles reflect the product at the time they were written.
- 5 min readAn open-source coding agent alternative: a guide to ShibaWhat the search for an open-source coding agent usually means, what Shiba actually is, and how to evaluate it against your own work.Read
- 5 min readWhy self-host a coding agent at allWhat self-hosted usually means, what it could mean, and the three belongings that decide which one you actually get.Read
- 4 min readApprovals in Slack, not another dashboardA Slack mention, a signed card, a named approver, and the run that only starts when the right person clicks.Read
- 4 min readWhat a failed run looks likeA run is stopped mid-flight, and the record it leaves behind is an error with evidence rather than a success badge.Read
- 4 min readHarness and model choices, and what they do not tell youThe situation is a per-run choice, the complication is that nobody has measured the outcome, and the resolution is to publish limits instead of claims.Read
- 4 min readCredentials that never enter the containerActors, keys, and handoffs traced from your machine to the sandbox and back out to the provider.Read
- 5 min readWhy the approval gate comes before the containerA planned run reaches an irreversible step, a human holds it, and the gate decides what happens next.Read