What a Healthy Agent Setup Looks Like
You don't need to build the loop — but you should recognize a well-run one. It has parts, guardrails, and instruments, and a team that can't point to all three is flying blind with something that takes actions.
A healthy agent has four things a leader can ask about. A loop — it acts, observes the result, and decides again. Tools — the specific, limited actions it's allowed to take. A stopping rule — a clear definition of "done" plus a hard budget so it can't run forever. And gates — human approval on anything irreversible. On top of that sit two instruments: an offline check that measures completion over a fixed set of tasks before you ship, and online monitoring plus traces that show what the agent actually did in production.
The single most important capability is the trace: a step-by-step record of every action the agent took and why. Without it, a failure is an unsolvable mystery; with it, you can see exactly which step went wrong. A team that can't show you a trace can't really operate an agent — they can only hope.
"Why did it do that?"
An agent took a baffling action in production and a customer is upset. You ask the team what happened. They can't say — there's no record of the agent's steps. What's the missing capability, and why is it non-negotiable?
The missing capability is tracing — a logged, step-by-step record of every action and decision. Without it, every incident is unexplainable, so you can't fix the cause, reassure the customer, or prove what happened. For something that takes real actions, that's not a nice-to-have; it's the difference between operating an agent and gambling with one. Require traces before you expand autonomy: if you can't reconstruct what it did, you can't be accountable for it.
OpsGrid: three weeks of quiet damage
OpsGrid ran an automation agent with an offline check but no live monitoring and thin logging. A routine change to one of its tools subtly altered its behavior, and the agent began mishandling a category of tasks — not crashing, just quietly doing them wrong. With no dashboard tracking completion and no usable traces, the pattern went unnoticed for three weeks, until a manager connected a run of complaints.
Reconstructing what had happened was nearly impossible because the traces weren't there. They added full step-by-step tracing and a live dashboard of completion and escalation rates; the next such regression was caught in hours, and diagnosed in minutes because the trace showed the exact failing step. Offline checks told them the tasks they'd anticipated still worked — only live monitoring and traces caught the one they hadn't.
Before launch, Tessera insists on two things: a full trace for every request the agent handles, and a dashboard showing completion, average steps, and escalations. The COO initially calls them overhead; both prove essential the first time an agent does something surprising.
Quiz · Chapter 2
- The four parts of a healthy agent are roughly:
- The single most important thing to operate an agent is:
- A healthy setup has:
- An agent with no hard budget: