AI That Ships October 29, 2025

The Secret to Useful AI Agents? Boring Tools.

Most “AI agent” demos overcomplicate things. When you’re starting out, keep it simple: one boring tool can beat a complex orchestration framework.

Last week I built a PubMed research agent in under an hour that runs multiple search strategies, filters 100+ papers against my criteria, and returns ranked results with evidence.

Zero custom code. Zero orchestration layers. Zero state management. Ask a question, get an answer.

I wrote a 3-line prompt, gave the agent an OpenAPI spec for the PubMed API (drafted by Claude from public docs), and enabled tool use. That was all.

The agent figured out the workflow: initial query, scan results, try synonyms, fetch more pages, synthesize, rank. It works like a human researcher—except it happily “reads” 100 abstracts while you’re still adjusting search terms.

Why it’s effective

Contracts beat instructions. OpenAPI tells the model what’s possible and how to do it.

Planning emerges naturally. Once the tool exists, the agent chains calls with minimal prompting (“try synonyms,” “expand acronyms,” “use author initials”).

Throughput beats precision. Instead of over-tuning a query, the agent makes smart decisions like “reading” 100 mediocre results and deciding what’s important.

Latency is cheap. Four API calls to refine a search equals 30 seconds of compute vs. hours of human review.

Try it today

You can prototype this in ChatGPT, though production belongs to cloud provider agents:

  1. Pick a read-only API you trust, such as PubMed, drug databases, or ClinicalTrials.gov.
  2. Ask Claude or ChatGPT to draft an OpenAPI 3.0 spec from the docs. Then, sanity-check types and auth.
  3. In ChatGPT, create a Custom GPT → Configure → Actions → Create New Action.
  4. Paste in your OpenAPI schema and supply auth (if needed).
  5. Write your prompt telling the agent about your specialized search needs. For example, you might say, “You are an expert in oncology biomarkers. Given the search terms, return only papers that discuss oncology biomarker topics…”

The cloud providers are unmatched for production, and you can run with extended planning and thinking. But you can graduate to that when needed.

Skip the Rube Goldberg web of agents in the beginning. An LLM + a clear API contract will give you big improvements.