Build it only if it earns its keep.

1. Should you build or buy?

Buy or subscribe when the job is common, the workflow is stable and a vendor already integrates with your stack. Build when the job is specific to your data, the volume makes per-seat pricing painful, or you need control over where data goes. A workflow with fixed steps is often better as a script than an agent.

Use the Agent Finder first. If three listed agents do not fit, that is a signal to build. If one fits, try it before writing code.

2. The four parts

• Model: pick the cheapest model that passes your evals. Do not start with the largest.
• Tools: each tool needs a clear name, typed inputs, a timeout and an error path. Fewer, better tools beat a giant toolbox.
• Memory: separate short-term conversation state from long-term facts. Store the minimum; make deletion possible.
• Loop: cap steps, cap spend and define what “done” means. Always have a human handoff for refunds, legal, medical or account deletion.

3. Pick a runtime

External links. We are not affiliated and do not take a commission.

4. A weekend build, in order

1. Write the job story. “When a support email arrives, classify it, draft a reply in our tone, and escalate refunds.” One paragraph.
2. Collect 20 real examples. Include the ugly ones. These become your eval set before you write prompts.
3. Build one tool. Connect the inbox read and the draft write. Make the agent stop before sending.
4. Add a trace. Log every model call, tool call, latency and cost. Without traces you cannot debug or price.
5. Add guardrails. Allow-list tools, cap steps at 8, cap spend per run, require approval for sends and payments.
6. Ship a boring UI. A queue with “draft ready / approve / edit” beats a chat box for most business workflows.

5. Keep it cheap

Cache repeated questions, route easy cases to a small model, batch classification, and set a monthly cap that pauses the agent rather than surprising you. Our own Phase 1 rules are the same: AI is used only where it earns its place, with a warning at 80% of budget and a stop at 100%.

6. Safety checklist

• Never let an agent send money, delete accounts or email customers without a human approval step.
• Treat web pages and tool output as untrusted input. Do not let them rewrite system instructions.
• Do not put secrets in prompts or Agent Cards.
• Log decisions, not just outputs, and keep a way to turn the agent off.

7. Get it listed

When it works, publish an Agent Card and list it free. A specific description and a public demo link are worth more than a launch thread.