# Give an agent a job. Give the system boundaries.

> Help people complete work that crosses documents and systems. Define what an AI agent may do, which decisions need approval and what happens when a step fails.

Canonical URL: https://marmation.com/solutions/ai-agents-and-workflows/

## The decision

Which steps can an assistant take without asking a person?

Separate reading, suggesting, approving and updating a record. Give each step specific permissions and a recovery path; allow more independent action only when tests justify it.

## A customer request needs interpretation and a recorded outcome

A customer request arrives by email with several documents. A team must classify the issue, check account context, prepare a response and update a case-management system. An AI application could reduce manual interpretation, but every action touches data, policy and customer consequences.

## An unattended demo can hide duplicate or unauthorized actions

A fully autonomous happy-path demo can conceal inconsistent tool responses, prompt injection, repeated transactions, missing identity context and vague ownership. A production workflow cannot assume the model will always choose the same tool for the same situation.

## Give each workflow step checks, permissions and recovery

Write the workflow as explicit steps: read permitted information, interpret the request, propose a response, obtain any required approval and save the result. Keep business checks in code and expose only the tool operations needed for each step. Define timeout and retry behavior, including duplicate prevention, so a lost response cannot create a second transaction. Give incomplete work a visible route to human review.

## The workflow, review and operating controls to agree

- Task/job charter and action-permission matrix for agent tools.
- Workflow engine or orchestration plan with deterministic gates and retry semantics.
- Review UI for uncertain or high-consequence outputs.
- Evaluation cases, transaction logs, rollback and operational monitoring and diagnostic records.


## Stable decisions may not need autonomous agents

A conventional rules engine is often better for stable structured decisions. A workflow orchestrator plus a small model call may outperform a sprawling multi-agent system. Autonomy should increase only where real evaluation evidence supports the change.

## Try hostile documents, repeated actions and unavailable systems

Test malicious content in retrieved documents, unexpected tool output, unauthorized requests, duplicate action replay, partial integration failure and safe human escalation. Follow current OWASP guidance and applicable security review.

## Build the exception path before expanding autonomy

Begin with a single workflow and one genuinely difficult step. Implement the exception path before promising that the system can run unattended.

## Related reading

- [Human approval is a system design decision, not a concession.](https://marmation.com/review/human-approval-in-ai-workflows/)
- [Threat-model the tools, not just the prompts.](https://marmation.com/review/threat-modeling-tool-using-ai-agents/)
- [Your systems may work. The workflow between them may not.](https://marmation.com/solutions/intelligent-business-systems/)