The demo-to-production gap in agentic AI.

The diagnosis

The pattern is consistent across the public record. Demos succeed by optimising one path: a polished prompt, a single tool call, a curated input. Production fails because the path is no longer single. Real users send malformed inputs, the network drops, the model retries, the queue backs up, the approval gate fires, the operator pauses the run, the session times out, the database write fails. Each of those is its own runtime concern, and each one is what the demo was not designed to handle. The work that closes the gap is not a better prompt. It is state, resumability, approvals, monitoring, and recovery: the architectural primitives operators have been building into systems for decades, applied to a new substrate. The principles below name what those primitives look like for agentic AI.

The numbers that name the gap

Every stat below links to its publisher. None are AI-summarised. Items the source research couldn't fully verify are flagged as such, not filled with vendor folklore.

Three commonly-asked numbers do not have a high-confidence public figure: average time-to-abandon for stalled agent projects, cost overrun rate versus initial estimate, and a universal eval-to-production regression rate. The validator beta is open to teams whose case studies would help close these gaps.

What practitioners say on the record

Every quote below is verbatim from the practitioner's own writing or talk. Each links to its primary source. None are paraphrased; none come from vendor marketing.

What the runtime ecosystem says it solves

This landscape reports each project's public claim alongside what the public record verifies that claim covers. Vendor inclusion implies neither endorsement nor competitive comparison. Items flagged in the source research as unverified are excluded; this list will expand as additional vendors are independently verified.

Five wrong moves the practitioner record names

These are the patterns named by the deep-research source as recurring misconceptions about closing the demo-to-production gap. Each card carries a counter-quote from the practitioner who named the limit publicly.

The doctrine answers each failure mode

Ten failure modes the source research surfaced as recurring on production agentic systems. For each, the AI Design Blueprint principles that address it.

1. 1

   Conversation as the only coordination model

   Multi-step work is driven through chat alone, with no underlying orchestration to recover, retry, or hand off cleanly.

   PrinciplesP4
2. 2

   Silent background failure

   Long-running work fails without surfacing progress or error to the operator who started the run.

   PrinciplesP2P3
3. 3

   Missing approval boundary

   Agents take consequential actions (writes, sends, payments) without an explicit human gate at the boundary.

   PrinciplesP7P8
4. 4

   State loss on refresh

   A page reload, tab close, or session timeout erases the agent's working memory mid-task.

   PrinciplesP6
5. 5

   Eval-to-production drift

   Pre-launch evaluations look clean; production traffic exposes a class of inputs the eval set never covered.

   PrinciplesP9
6. 6

   Tool-use cascade hallucination

   An agent invents tool arguments or response shapes; the next step trusts the fabrication and acts on it.

   PrinciplesP4P7
7. 7

   Context window collapse

   Long conversations exceed the model's window; older context drops silently and behaviour degrades without warning.

   PrinciplesP6
8. 8

   Approval gate decoration

   An approval prompt exists, but the operator can only click 'approve': no inspect, no edit, no rejection path with a reason.

   PrinciplesP7P8
9. 9

   Demo-only happy path

   The system works on the scripted demo flow; the second branch is never built or even named.

   PrinciplesP9P10
10. 10

    Loop detection failure

    An agent retries the same failing action indefinitely, burning tokens and time without escalating to a human.

    PrinciplesP3P10