The Model Is the Easy Part: Why Production AI Agents Break on State

When an agent pilot stalls, the first instinct is to try a better model. In production, that is rarely where the problem is.

Ido Fishman, who has run AI agents in production for six months, put it plainly in a recent HackerNoon post: building state "turned out to be about 70 percent of the work, and the thing that broke most often." His first agent took three weeks. His fourth took two hours, because by then the plumbing existed. That is one builder's experience, but it matches what we see inside client systems.

What state actually means

Most demos keep chat history: a record of what was said. Production needs a record of what is true now. Which supplier negotiations are open, what was approved last week, which payment already went out, and what the next step is. Chat history can tell you the agent "sent a follow-up". It can't reliably tell you whether the supplier replied, or whether the agent already paid them.

Three failures no better model fixes

• The repeat. A call times out, the agent retries, and the email or payment goes out twice. The fix is idempotent writes: every action carries an ID, so a retry can't do it again.
• The restart. A nine-step job fails at step six and starts again from step one, redoing work and sometimes actions. The fix is checkpoints, so it resumes where it stopped.
• The silence. The agent stops doing its job and nothing tells anyone. SaaStr, which runs more than 20 agents with three people, found one had "silently failed to upload new content for four months." The fix is outcome checks: alerts when expected work stops arriving.

Silent failures also cost money. Fishman describes an agent that went from about €2 to about €40 a day, stuck retrying a failing tool call, and nobody noticed for a while.

Six things to build before the second agent

• A durable task record: status, owner and next step for every job, kept outside the conversation.
• Idempotent writes: a retry can never send, pay or delete twice.
• Checkpoints: a crash at step six resumes at step six.
• A log of every run: inputs, tool calls, results and final actions.
• Outcome checks: an alert when expected work stops happening.
• A daily cost check per agent: a loop shows up in a day, not on the invoice.

None of this depends on which model you use. All of it decides whether your agent can run for months without someone watching it.

How we approach it at Zylen

Zylen is an AI optimization partner. We work inside the AI systems companies already run and improve cost, security and data control one measured change at a time. Our ten-day AI Optimization Audit finds where your agents lose state, repeat work or fail silently, and what to fix first.

Fix what's below the waterline: start the AI Optimization Audit →

Sources: Ido Fishman, "Six Months of Running AI Agents Taught Me the Model Is the Easy Part", HackerNoon, 8 Oct 2026; SaaStr, "Introducing The Agents".