First-party system · AI operations

An operating system for people and AI agents

A real first-party incident exposed the difference between a plausible plan, executable work and a confirmed outcome.

Innokenty Bodrov · Product lead

← All work

01 / The task

I run a portfolio of products and publishing workflows with several AI agents. Written roles existed, but different processes could act from stale versions of the same plan. That created work-item collision risk and a gap between editorial intent and the live execution queue.

02 / What happened

A readback on 28 September 2026 found a candidate editorial plan but zero tasks due in the live queue and zero confirmed publication facts. The plan had not become an executable slot with an owner, accepted revision, destination and release path.

A separate channel review found an X adaptation duplicating the primary Threads placement and a Dzen item repeating the same evidence. Both duplicate placements were cancelled; Threads remained the sole primary route for that item.

03 / My correction

I reconciled plans against live records and separated three kinds of fact: the plan records intent, a task records executable work, and a publication record confirms what went live. A ready item now needs a named owner, exact accepted revision, destination, evidence and execution path.

I also specified a stronger technical control plane with version checks and work leases. That design is a pilot candidate, not a deployed control claimed by this case.

What this does not prove

The collisions were reconciled without final data loss and redundant placements were removed. No throughput gain, client ROI, conversion uplift or measured reduction in incident frequency is claimed. This is first-party operating evidence, not a client case.

Source and verification

This account is based on my internal decision record from 28 September 2026. The underlying queue and record are not public; the linked earlier publication is also my own account.

Earlier author publication ↗

Working on a similar problem?

Tell me about the workflow and the decision your team needs to make. I will read your note myself.

Discuss AI operating rules
Back to top