AI Agent GA — Eight Months of Decisions Before Launch

CollabOps AI Agent went GA yesterday. Eight months after beta start. The original plan was three months of beta. The reason it became eight: four launch holds.

This post is those four decisions and why holding was cheaper than launching.

Hold 1 — insufficient permission model

Three months in. All technical features ready. Working at 12 beta customers. Launch was scheduled.

That moment, security found in internal review — agent actions logged to a different audit trail than human actions. To find what an agent did, you'd have to query a separate system after the fact.

That gap guaranteed failure in customer security reviews. Launch held two months to integrate the permission model onto the graph (→ permission model redesign).

Hold 2 — on-prem deployment mode

Five months in. After the permission integration, launch was scheduled again. Then the first on-prem PoC broke a model-hosting assumption. We had assumed cloud LLM calls; that customer required every model call inside their network.

This wasn't a feature add. It was an architectural change. Launch held two months — self-hosted model option + internal MCP server deployment → MCP guide.

Hold 3 — non-determinism

Seven months in. After two holds, launch was scheduled again. A beta customer reported different outputs for the same input. Their security team concluded — an agent with production authority producing different actions for the same input won't pass audit.

Fix: wrap the decision step of LLM calls in a deterministic rule engine. The model proposes candidate actions; rules choose. Launch held three weeks.

Hold 4 — docs / SDK consistency

7.5 months in. After three holds, almost launching. A final pass found — API docs, CLI, and SDK expressed three different permission models. CLI showed a simple model; API showed the graph model; SDK showed a transitional model. Users got different mental models depending on entry point.

Fix: unified all three on the graph model. Launch held three weeks. The smallest hold — and the most clearly correct.

Eight months — the cost

Planned          | Actual          | Diff
─────────────────┼─────────────────┼────────────
Beta: 3 months   | Beta: 8 months  | +5 months
Launch work: 6mo | Launch work: 8mo| +2 months
Expected post-launch incidents | Actual: 0 | −3 to −5

Five extra months of beta turned an expected 3–5 incidents into zero. Each post-launch incident costs 1–2 months of fallout (urgent patches, re-running security review, rebuilding trust). Net: holding five months made the launch faster overall.

Launch is the start, not the end

GA isn't the end of decisions. Next ones already queued:

• Authority boundary expansion — how far in the next six months
• Multi-agent vs single-agent — our position is multi-agent is almost never needed
• Per-customer fine-tuning option — our standard model vs customer-domain fine-tune

Technical readiness is not a launch signal

For an AI agent product, "all technical features are ready" is not a launch signal. Our four holds — permission model, on-prem deployment, non-determinism, docs/SDK consistency — were all found after the features worked. If your team is preparing an agent launch, answer the four hold reasons for yourselves. If three or more answers are fuzzy, another hold is cheaper than launching. Five extra months of holds turned our expected 3–5 incidents into zero.