Your agent is not broken. Your platform was never built to run non-deterministic work.

Gartner has placed 40% of enterprise agentic AI initiatives at cancellation risk by 2027. Industry reporting puts the 'never reaches production' rate higher than that — closer to 88%. Every CTO I have spoken to in the last six months has an agent initiative on the board roadmap. Almost none of them are running one in production yet.

The conversation about why is dominated by the wrong actors. AI tooling vendors blame the model. Management consultants blame the use case. The engineers building the agent know the truth and rarely get to say it out loud.

An HTTP service has a known production contract. You give it CPU and memory limits, you give it a rate limit, you give it a latency SLO, you give it a circuit breaker, you give it a deployment pipeline, you give it a rollback procedure. The contract is so well understood that platform teams build entire abstractions on top of it.

An agent has none of that. Agents are non-deterministic. They make tool calls. They loop. They consume tokens at variable rates. They can stay alive for minutes or hours. Their output is not just a response — it is a sequence of side effects on internal systems. The platform underneath them was built on the assumption that the workload returns in a few hundred milliseconds and is stateless. The agent is neither.

Here is the agent production contract your platform needs and almost certainly does not have.

• Token budget per run. Not per request. A single agent run can consume 10 to 100 times what a single LLM call costs. Without a budget at the run level, one stuck tool call can burn a month of FinOps headroom in an hour.
• Max tool iterations. A circuit breaker on agent loops. Without it, a malformed tool response triggers retry storms that look like progress until the bill arrives.
• Eval floor. A continuous evaluation pipeline that fails the deploy if output quality drops below a threshold. Not just CI. Production gate.
• Tool surface provenance. Which tools is the agent allowed to call? Who validated them? What is their failure mode? An agent with broad tool access has the same blast radius as a service account with admin permissions.
• Gateway SLO. The AI gateway is now your highest-blast-radius internal service. It is between every AI feature in the company and the model providers. Did anyone give it an SLO, a failover region, and a named owner?
• Non-deterministic rollback. Rolling back an agent service is not the same as rolling back a stateless API. Embeddings, stored decisions, and intermediate state from the new version persist after the rollback.

Most agent initiatives fail to reach production because they cannot satisfy any of these six. The team building the agent does not have authority over the platform. The platform team does not understand the runtime characteristics of the workload. The CTO is told it is a model problem. It is not. It is a contract problem, and the contract has not been written.

If your agent initiative has stalled for more than six months, the question to ask is not 'is the model working?' The question is 'what is the production contract for this workload, and which of the six clauses does my platform fail?' Until that contract exists and is enforced at the platform layer, the agent will not run in production. Adding more prompts will not save it. Hiring more researchers will not save it.