Use a chat for bounded one-off thinking, one agent for repeatable multi-step work involving tools, and a team only when parallel specialist work materially improves the result. Keep high-stakes final decisions human-led.
When a normal chat is enough
A chat is often the best interface when the user can stay present, provide the relevant material, evaluate each answer, and correct the system immediately. Examples include drafting from notes, summarizing one document, brainstorming names, explaining a spreadsheet, or outlining a plan.
Do not build an agent merely because a prompt has several steps. A well-structured chat can follow a checklist, ask questions, and revise its work while preserving the cheapest and clearest form of human oversight.
When one agent makes sense
An agent is useful when a repeated goal requires multiple actions, external tools, files, memory, monitoring, or a schedule. A weekly reporting workflow might retrieve files, calculate changes, draft a summary, save the result, and wait for approval.
A dependable agent needs more than a goal. Define its trigger, permitted tools, data access, output format, stop conditions, error handling, logs, and the point at which a human must approve the next action.
When an agent team is justified
A team can help when several workstreams are genuinely independent and can run in parallel: regional market research, separate technical reviews, multiple source audits, or specialist analyses followed by synthesis.
The team needs a coordinator that assigns non-overlapping work, a shared evidence format, a process for resolving contradictions, and a final reviewer. Without these, multiple agents can produce five polished fragments rather than one trustworthy answer.
When the answer is none of the above
AI can assist with preparation but should not own final medical diagnoses, legal judgments, employment decisions, credit decisions, safety-critical actions, or irreversible financial commitments. These require accountable human judgment and, often, licensed expertise.
“None” also applies when the task is physical, the system lacks necessary access, success cannot be defined, evidence is unavailable, or the cost of supervising the automation exceeds the work saved.
A seven-question architecture test
- Is the task one-off or repeated?
- Can success be stated in observable terms?
- Does it need files, web access, email, calendar, code, or another tool?
- Are there independent workstreams that can truly run in parallel?
- Can a human review the output before anything consequential happens?
- What data and permissions are required?
- What happens when the model is wrong, the tool fails, or the evidence conflicts?
Why agent teams cost more than they appear
A team multiplies model calls. Each specialist receives instructions and context, may call tools, may retry, and then sends findings to a coordinator. The coordinator consumes those findings, resolves inconsistencies, and often asks for revisions. A workflow described as “five agents” can create dozens of calls.
Token cost is only one burden. There is also latency, debugging, permission management, duplicate work, conflicting conclusions, and the time required to inspect the final result.
How to build safely
- Pilot the hardest representative case. Do not validate only on an easy demo.
- Use least privilege. Read access does not imply write access; drafting does not imply publishing.
- Make consequential actions reversible. Save drafts before sending, stage changes before committing, and require approval before moving money.
- Preserve evidence. Ask for source links, quoted passages, calculations, and timestamps.
- Set stop conditions. Limit loops, retries, spend, and elapsed time.
- Measure net value. Include supervision and correction time, not just execution time.
Examples
| Task | Best starting point | Why |
|---|---|---|
| Rewrite a customer announcement | Chat | Bounded, one-off, and easy to review. |
| Prepare a weekly sales summary from approved files | One agent | Repeated, multi-step, tool-using work with a clear output. |
| Research five markets and synthesize a cited report | Small agent team | Independent parallel research can be coordinated and reviewed. |
| Decide which employee to terminate | Human-led | High-stakes employment judgment; AI may organize evidence but should not decide. |