Agents and Workflows
Choose autonomy only where adapting the next step creates value. · AI Product Management · Lesson 49 · 3 min
Agents and Workflows · 3 min
Situation
A simple task gets an elaborate agent.
A team wants to turn a support ticket into a category, suggested reply, and review task. Someone proposes an autonomous agent that can search, plan, call tools, and repeat until done.
The extra freedom may be useful, but a known sequence might solve this workflow with fewer failure modes.
Mental model
Three degrees of control.
A fixed workflow follows predefined steps and rules. An LLM-assisted workflow uses a model for selected steps while software controls the sequence. An agentic loop lets the model choose subsequent actions based on observations toward a goal.
These form a spectrum. “Agent” is not a guarantee of quality or a requirement for a useful AI product.
Example
Choose the simplest adequate path.
For ticket triage, software can fetch the ticket, ask a model for a structured category and draft, then send it to a human review queue. The sequence is predictable.
For an investigation where each result determines the next source to inspect, a bounded agentic loop may provide more value. The task's uncertainty justifies adaptation.
Failure case
Each additional step creates another way to fail.
An agent can select the wrong tool, repeat work, lose context, spend too much, or interpret an error as success. Even individually reliable steps can accumulate failure across a long run.
Give the loop limits on time, spending, tools, scope, and repeated failures. Define a stopping condition and a visible account of confirmed outcomes.
PM decision
Specify the permitted autonomy.
Ask which decisions the model may make, which actions need approval, and which operations are impossible by design. Keep irreversible or sensitive actions behind appropriate controls.
Evaluate complete tasks, not only isolated model answers. Include recovery after tool failure and cases where the right behavior is to stop and ask for help.
Remember this
Autonomy is a trade-off, not a maturity level.
Use a fixed path when the process is known. Add adaptive action selection where it improves outcomes enough to justify the cost of monitoring and failure handling.