Your AI Agent Doesn't Need More Autonomy. It Needs a Process.

There's a bad habit running through a lot of AI agent builds right now. Someone gets excited about what a model can do. They give it fifteen tools, a wall of instructions, and a vague directive to "handle the workflow." The agent loops, reasons, retries, burns tokens, occasionally does the right thing , and everyone calls it progress. It's not progress. It's expensive improvisation. Google just shipped ADK 2.0, and buried inside the launch post is something that should reset how most businesses think about agents. Moving from prototype to production, they found, means confronting a different class of problems: unpredictable loops, hallucinated business logic, context bloat, messy error handling, and execution you can't inspect or trust. Their answer wasn't "get a better model." It was workflows. Graph-based control. Deterministic routing. Human approval. Retry policies. Logs. Typed steps. Observability. In other words: software.

The received wisdom AI agents are the new employees. Give them enough tools and autonomy, and they'll figure out the process. The model is the intelligence , let it decide.

This isn't wrong in the way that a lie is wrong. It's wrong in the way that a bad habit is wrong. It works just often enough in demos that people keep doing it. In production, it breaks. Reliably.

Google's refund example in the ADK post is perfect precisely because it's boring. A refund needs to check purchase history, read policy, assess eligibility, issue money, email the customer, and close the ticket. Some of those steps need judgement. Most of them don't. Checking a purchase record doesn't need creativity. Routing a completed case to "closed" doesn't need a model to have a little think. And issuing money should not be left to a probabilistic loop because the system prompt said "be careful with transactions." But interpreting a customer's complaint? Classifying ambiguous intent? Drafting an empathetic response that reflects policy without sounding like legal boilerplate? That's exactly where a language model earns its cost. The rule is simple. If step B always follows step A, don't pay a language model to decide it. Put it in the workflow. Save the model for the parts that actually need judgement.

What this looks like in practice Take any repeatable business process , lead triage, support routing, content production, quote drafting , and split it into four buckets:

Deterministic steps. These belong in code. Fetch the record. Apply the rule. Check the field. Update the status. Normal software handles this faster, cheaper, and with zero hallucination risk. Judgement steps. These are good AI territory. Interpret a messy message. Classify intent. Summarise evidence. Draft a reply. Spot a mismatch between what the customer said and what the policy allows. Approval steps. These need a human in the loop, at least until the system has earned trust. Money moving. Legal exposure. Deleting data. Publishing externally. Brand-sensitive decisions. Keep a human here. Measurement steps. The ones people forget because they're not flashy. What happened? What did the agent use? Where did it fail? What did the human change? Did this actually save time? Did quality go up? Did something risky quietly slide through? That last bucket is where most AI projects die. The agent runs. Nobody measures. Nobody knows if it's working. It becomes a liability disguised as a feature.

The timing matters here. Anthropic just launched Claude Sonnet 5 , more capable on tool use, better at follow-through, cheaper than the bigger models. The pattern that keeps coming back: models are getting better, cheaper, and more autonomous-feeling. But a cheaper, more capable model doesn't remove the need for process. It increases how far the system can get before you notice something's wrong. More autonomy in a badly-designed loop doesn't help. It just means the agent confidently reaches further down the wrong path before anyone catches it. The live conversation on X right now echoes this. Last week someone put it bluntly: "Should this even be an AI agent? Companies are replacing simple automation with expensive AI agents. Sometimes that's the right decision. Often it isn't." Another engineer made the same point from a production angle: a workflow is a process where the output is deterministic or waiting for human review. An AI agent shouldn't be processing a workflow in production. These aren't contrarians. These are the people running real systems.

Where the agency model changes

The old agency sold output , pages, campaigns, ads, reports. The AI-era trap is selling more output, faster. More posts. More variants. More hollow personalisation. More content that looks right and means nothing. The useful model is different. It sells operating capacity. That means building the loop: research in, draft out, claim checks, approval, scheduling, performance review, lessons fed back into the next brief. A system that creates the work, verifies the work, routes the work, and learns from what happened. Not "we built you an agent." "We removed the part of your process that was burning three hours a week, and we can prove it."

The businesses that win the next phase of AI won't be the ones with the most autonomous agents. They'll be the ones that did the boring work first. Mapped the real workflow. Decided the approval rules. Built the eval sets. Put logs somewhere humans can read. Measured whether anything actually improved. That's not a lack of ambition. That's how you get agents out of the toy box and into the business without pretending probability is process. Autonomy is not the product. Reliable work removed is.

→ Foundry Works builds AI operating layers for growing businesses. Talk to us about what that looks like for your team.

Word count: ~820

X Post Copy

Most AI agents are failing for the same reason: no one mapped the actual workflow first. If step B always follows step A, don't pay a language model to decide it. Google's ADK 2.0 said it plainly. Production agents need workflows, approval, logs, and software doing the boring parts. foundryworks.ai/blog/

LinkedIn Teaser

There's a bad habit running through a lot of AI agent builds right now. Someone gets excited about what a model can do. They give it fifteen tools, a wall of instructions, and a vague directive to "handle the workflow." Then everyone acts surprised when it loops, burns tokens, and occasionally wanders off in the wrong direction. The rule that fixes most of this: if step B always follows step A, don't pay a language model to decide it. Put it in the workflow. Save the model for the parts that actually need judgement. Read the full piece → foundryworks.ai/blog/your-agent-doesnt-need-more-autonomy

Further reading

• The Agent Problem Isn't Building Them
• Your Company Isn't Documented Enough for AI Agents
• Your Website Needs a Tool Surface