Anthropic's Agent Decision Checklist: 4 Questions to Determine If You Should Build an Agent

In May 2026, an internal presentation from Anthropic was leaked, featuring an incredibly insightful decision-making checklist on a slide: the "Should I build an agent" Checklist.

My first thought upon seeing it: This one diagram could save 90% of AI founders six months of wasted effort.

Because the most common mistake right now is—trying to turn every single need into an Agent. The result is burning through cash on a product that still can't run reliably.

This article breaks down that checklist, giving you a practical decision-making framework you can use right away.

---

1. Why is this checklist important?

Over the past year, the biggest misconception in the AI world has been treating "Agent" as a universal solution.

The boss asks: Can we build an Agent? The product manager asks: Should we implement an Agent? The investor asks: Have you 'Agent-ified' your product?

But what Anthropic wants to tell you is: An Agent is an "expensive thing"; not every task deserves one.

Using it in the wrong place is just burning money.

---

2. Anthropic's 4 Questions

1. Is the task complex enough?

The criteria are very straightforward:

• Single-task value < $0.1 → Use a Workflow (hard-coded process)
• Single-task value > $1 → Use an Agent (let the model make its own decisions)

Why? Each Agent call consumes a massive number of tokens for planning, reflection, and tool use. If the task itself isn't valuable, the Agent's overhead will eat up all your profit.

2. Can every step of the task be performed?

This is the most easily overlooked point.

• If there's any step the Agent cannot do → First, narrow the scope
• Only if the entire process can be executed → Then consider building an Agent

Why it's crucial: An Agent is an end-to-end commitment. If it gets stuck at any intermediate step, the entire chain breaks. It's better to first solidify the "doable parts."

3. Is the cost of error high?

• High cost → Make it read-only or implement human-in-the-loop review
• Low cost → Only then let the Agent run autonomously

For example:

• An Agent drafts an email for you (requiring you to click send) → Safe
• An Agent directly sends an email to a client → If it says the wrong thing, the company's reputation is on the line

4. Can errors be detected promptly?

This point complements point 3:

• Errors go undetected → Dangerous, don't deploy an Agent
• Errors are immediately visible → Can proceed

Core logic: Agents will make mistakes. The key is whether you can catch and handle them when they do.

---

3. Summarizing the checklist in one sentence

Agents are suitable for: High-value, fully controllable, fault-tolerant tasks where errors are observable.

Conversely, if any one condition is not met, prioritize a Workflow; don't force an Agent.

---

4. The 3 most common pitfalls

Pitfall 1: Building an Agent for a low-frequency task

Forcing a multi-turn conversational Agent for an internal tool used 10 times a day. Result: Development costs are 5x that of a Workflow, and performance is worse.

Pitfall 2: Equipping an Agent with tools it can't use

An Agent without database permissions is tasked to "automatically analyze sales data." Result: The Agent fabricates data, misleading the boss.

Pitfall 3: Running an Agent where no one is watching

Running an Agent in a backend cron job with no logs or alerts. Result: It crashes for three months unnoticed, while the bills keep coming.

---

5. So, what areas are currently suitable for Agents?

Filtering by Anthropic's criteria, there are actually not many scenarios truly ripe for Agent-ification right now:

• Programming assistance (high task value, errors immediately visible) → Cursor / Claude Code
• Customer support (high task value, human oversight as a safety net) → Intercom Fin
• Data analysis (complex tasks, human review of outputs) → Hex / Julius
• Operations automation (long processes, can be reviewed step-by-step) → Lindy / Relay

Commonality: Either the task is expensive enough, or errors are easy enough to catch and handle.

---

6. One piece of advice for founders

Don't ask "Can I build an Agent?" First ask, "Is this task worth $1?"

If yes, then consider an Agent. If no, stick diligently to a Workflow. Get the process running, running stably, and running profitably.

An Agent isn't better because it's more complex or more "expensive."

This is the unspoken message Anthropic is sending to all AI founders with this checklist.

---

Final word

If you're building an AI product, post these 4 questions at your workstation:

1. Is the task valuable enough?
2. Can every step be performed?
3. Is the cost of error high?
4. Can errors be detected?

Only proceed with an Agent if you get 4 green lights.

This is the most pragmatic decision-making standard for building AI products in 2026.

References

1. Building Effective AI Agents — Anthropic — A seminal post from Anthropic's official engineering blog, systematically discussing the trade-offs between Workflows and Agents. The source of the decision-making framework in this article.