When You Actually Need an Agent
"Agent" is the word of the moment, so teams reach for one everywhere. But an agent is the right tool only when the work genuinely needs a loop. Using one where a simple script would do is how you turn a reliable process into an unpredictable, expensive one.
Here's the test. If a task follows a fixed sequence of steps every time — do A, then B, then C — a plain script or a single AI call is cheaper, faster, and far more reliable, because it does the same thing every time. You need an agent only when the path can't be scripted in advance: when the next step depends on what the last step returned, when the task branches unpredictably, when it takes several tries to get right. The loop is the cost you pay for that flexibility, and it buys you nothing on work that was never variable.
As a leader, the failure to watch for is "agent-washing" — wrapping a deterministic workflow in an agent because it sounds advanced. You inherit all of an agent's risks (nondeterminism, cost, wrong actions) and none of its benefits. The right question isn't "can we make this an agent?" — it's "does this task actually need a loop?"
Agent or script?
Your team wants to build an agent to "process incoming invoices": extract fields, validate, enter into the system. It's the same three steps every time. Is an agent the right call?
Probably not. If it's the same fixed sequence every time — extract, validate, enter — that's a scripted workflow (possibly with one AI call for the extraction), which will be cheaper, faster, and more reliable than a loop. Reserve the agent for the genuinely variable parts: an invoice that doesn't match, a vendor query that needs investigation, an exception that branches. Ask the team, "what about this task can't be written as fixed steps?" If the honest answer is "nothing," you don't need an agent — you need a script with maybe one smart step.
Straightline: an agent for a straight line
Straightline rebuilt a stable, deterministic onboarding workflow — create account, send welcome, provision access, notify manager — as an autonomous agent, because leadership wanted "AI-powered onboarding." The result was worse on every axis: it occasionally did steps out of order, sometimes skipped one, cost far more per run than the old script, and was harder to debug because each run could differ. They'd taken a process that worked 100% of the time and made it work 90%, expensively.
They reverted the fixed parts to a plain workflow and kept an AI step only where judgment was actually needed — handling unusual access requests that didn't fit the template. Reliability went back to near-perfect and cost dropped. The lesson for their VP: an agent is for the parts that vary; forcing it onto a straight-line process just adds nondeterminism you didn't want.
Tessera scopes its agent carefully: routine, fixed requests (password resets, standard access grants) stay as scripts, and the agent handles only the messy, multi-step requests where the path depends on what it finds. That scoping keeps the reliable stuff reliable.
Quiz · Chapter 3
- You need an agent (a loop) when:
- Using an agent for a fixed workflow:
- "Agent-washing" means:
- The right question is: