Skip to content
← All articles
5 min readAgentic AI workflow design

Designing agentic systems with useful boundaries

Agentic AI becomes easier to engineer when the workflow, tool permissions, and human approval points are designed before the model is asked to act.

Start with one workflow, not an agent

“Agent” is a vague unit of scope. A workflow is something you can describe: a trigger, a sequence of steps, a decision, a handoff, and a definition of done. Start there. Pick a workflow where the current process is repetitive, bounded, and observable, and where a wrong action is recoverable.

This framing keeps the project honest. It gives you something to evaluate, a place to insert approvals, and a boundary for tool access. If the workflow cannot be described without inventing a new organization, it is probably not the first one to automate.

Choose the smallest model and the narrowest context that clear the evaluation. Larger models can be useful for difficult cases, but they also raise latency, cost, and the number of ways the workflow can surprise you. Start with a model that is good enough for the bounded workflow, measure it, and escalate deliberately where the evidence supports it.

Separate suggestions from actions

A model can often recommend a next step without being allowed to execute it. That separation is one of the simplest safety properties available. Suggestions can be reviewed, logged, and compared. Actions change state. Start with read-only tools and a human who performs the write. Move the write behind the agent only when the evaluation shows the recommendation is reliable and the action is reversible.

This is not a permanent limitation. It is a sequence that lets the team learn where the model is dependable before granting it consequences.

Tools are permissions

Every tool is an access decision. A search tool reads data. A ticketing tool writes it. A deployment tool changes production. The tool schema is also an interface constraint: narrow inputs produce more predictable behavior than a general “run command” tool.

Define each tool with the smallest useful scope, an explicit identity, and an audit trail. Separate read and write tools even when they hit the same system. Make destructive or irreversible actions require a distinct approval step, not just a different parameter. Assume the model will eventually call a tool in an unexpected order. The tool, not the prompt, should enforce the boundary.

State and memory are data decisions

Agentic systems accumulate context: conversation, retrieved documents, intermediate results, and memory across sessions. Each is a data-handling decision with residency, retention, access, and deletion implications. A memory feature that stores customer data indefinitely is a compliance problem wearing a product label.

Decide what the workflow needs to remember, for how long, and who can read it. Keep retrieved context attributable to its source. Prefer reconstructing state from systems of record over letting the agent maintain a private copy. If memory is necessary, make it inspectable and deletable, and scope it to the user or case rather than the whole tenant.

Evaluation is a product requirement

A demo succeeds on a good example. A workflow has to succeed on the boring and adversarial ones. Build an evaluation set from real cases: common requests, ambiguous requests, missing data, conflicting instructions, and cases where the right answer is to escalate or refuse.

Evaluate the whole workflow, not just the final text: tool choice, arguments, approval behavior, latency, and cost. Keep a small regression set and run it when the prompt, model, or tools change. Record failures with enough context to reproduce them. Without this, every improvement is a guess and every regression is discovered by a user.

Version the things that change behavior: prompts, tool definitions, model choice, and retrieval configuration. A workflow that cannot say which version produced a bad outcome cannot be debugged or rolled back. Keep the evaluation set stable enough to compare versions, and refresh it when the workflow’s real inputs change. Evaluation is not a one-time gate. It is the feedback loop that makes the system improvable.

Human approval belongs in the design

Approval is not a modal dialog bolted on at the end. It is a step with an owner, a decision, and a record. Decide which actions require approval, who can give it, what context they see, and what happens when they are unavailable. A reviewer who sees only “approve?” will approve.

Give the reviewer the proposed action, the evidence behind it, the affected systems, and the reversibility. Make rejection a normal path that returns useful information to the workflow. For high-volume workflows, design escalation tiers so that routine, reversible actions do not consume the same attention as consequential ones.

From prototype to operable system

The gap between a compelling prototype and an operable system is mostly engineering discipline: contracts between components, failure handling, idempotency, telemetry, cost controls, and ownership. A workflow that works when every dependency is healthy is a prototype. An operable workflow has a plan for a slow model, a failed tool, a duplicated request, and a mistaken action.

Before calling it production-ready, answer six questions. Who owns it? How is it monitored? What does it cost per case? How are prompts and tools versioned? How are incidents reviewed? How is access revoked? Those answers are less exciting than the demo and more predictive of whether the system survives contact with real work.

Incident response needs its own design. Decide how a bad action is detected, who can stop the workflow, how state is reconciled after a partial failure, and how affected users are told. A kill switch is necessary but not sufficient. The team also needs a way to replay or reverse the action safely. Write the runbook before the first production case, not after the first incident.

---

Related service

Agentic AI workflow design

Next step

If this is the decision in front of you, describe the constraint. I will tell you whether the work fits.

Start a project inquiry