What We Look for Before Recommending an AI Agent

Autonomous agents are the right tool for a narrower set of problems than the marketing suggests. Before we propose an agentic workflow over a simpler deterministic pipeline, we run through a short checklist — and more often than we'd like to admit, the answer is 'build the boring version instead.'

The three questions we ask first

• Does the task genuinely require multi-step, model-directed decisions — or could a fixed sequence of steps with a model in the loop handle it just as well?
• Can errors be caught and recovered from cheaply, or does a wrong decision compound silently across several steps before anyone notices?
• Does the value of automating this task justify the added unpredictability of letting a model choose its own path?

If any answer points the wrong way, a deterministic workflow with a model handling the genuinely ambiguous step usually wins — it's cheaper to build, far easier to debug, and dramatically easier to explain to a stakeholder asking why the system did what it did.

Where agents earn their complexity

Agents pay for their added complexity on tasks that are open-ended by nature: multi-system research where the next step depends entirely on what the previous step turned up, orchestration across several internal tools where the right order can't be fully enumerated in advance, or long-running investigative work where a human would naturally improvise. We've shipped agentic systems for exactly those cases, with real guardrails — scoped tool access, hard step limits, and a human checkpoint before anything irreversible happens.

For everything else — and that's most of what crosses our desk — we'd rather ship something boring that works, monitor it for a quarter, and only reach for an agent once we've proven the simpler version genuinely can't keep up.