The Anatomy of a Reliable Agent
People ask which model is most reliable. It is the wrong question, in the same way that asking which engine is most reliable says nothing about whether the car will start.
Reliability in an agent comes from five parts. One of them is the model.
A job, not a chat
An agent with a job description knows what it owns: an outcome, a trigger, and a definition of done. An agent without one is a chat window that occasionally does something.
Write the job as a sentence a sceptical colleague could agree with. If you cannot, the agent will drift, because there is nothing to drift away from.
A boundary
What it may decide alone, what it must escalate, what it must never touch.
The escalation path matters more than people expect. An agent that hands over badly is worse than no agent: a human receives a task with no context and now has to reconstruct what happened. Hand over the transcript, the attempts, and what the agent thinks the problem is.
Tools with contracts
Each tool the agent can call needs a narrow interface, typed inputs, and a failure mode that makes sense to a model. A tool that returns a 300-line JSON blob on error teaches the agent nothing. A tool that returns "no record found for that reference" teaches it to try a different reference.
Also: the smallest set of tools that does the job. Every tool is a way for the agent to be wrong in a new way.
Memory that is explicit
Three kinds, and conflating them causes most of the confusion:
- In-conversation: what has been said so far. Short, cheap, automatic.
- Across sessions: what we learned about this customer. Needs a write policy and a way to be wrong and corrected.
- Domain knowledge: belongs in retrieval, not in memory. If the fact is true for everyone, it is a document, not a memory.
A way to be wrong quietly
Refusal, escalation, and a human in the loop on the expensive paths.
The measure of a reliable agent is not that it never fails. It is that when it fails, it fails in a way that is visible, bounded, and recoverable.
Why the model is the least of it
Swap a weaker model into a well-built agent and it gets a bit worse at the edges. Swap a stronger model into a badly-built one and it becomes a faster, more fluent version of the same failure.
Build the five parts. Then the model choice is a cost decision, which is what it should have been all along.
Those five parts are the whole of an AI agent build. The model is the last thing chosen, not the first.