Build a research agent
you can check.

A research brief is only useful if you can trace its claims. Define the subject, source policy, and result format before asking the agent to search. Do not use the length of the final brief as a measure of quality.

Start with a small source set.

The downloadable example reviews source notes supplied as input. It does not browse or verify live publication dates. This first step lets you test whether claims stay within the evidence before adding web access.

Download the research method and sample inputs. The sample uses a clearly labelled fictional lab note. It is test data, not a real research finding.

curl -fL https://withmethod.ai/examples/research-brief.method -o research-brief.method
curl -fL https://withmethod.ai/examples/research-inputs.json -o research-inputs.json
method validate research-brief.method
method run research-brief.method --inputs research-inputs.json --workspace . --run-dir .runs/research

Install and check the SDK first with the quickstart. The example writes a local Markdown file. It does not send a message or publish a report.

Use two steps with separate checks.

1. Extract findings. Read the supplied notes. Keep only findings relevant to the requested topic. Record the source ID and an exact supporting quotation for each claim.
2. Write the brief. Use only the checked findings. Preserve source IDs, separate facts from limits, and say when no supported findings remain.

The first check compares each quotation and claim with the input notes. The second compares the brief with the extracted findings. These are model checks, so review the saved evidence yourself before using the result.

Know what a good result looks like.

With the supplied fictional note, an acceptable brief would say that a simulated gripper completed 8 of 10 trials. It would also state that there were no hardware trials. It must not turn that result into a claim about real-world robot reliability.

Change the input to an unrelated topic. The method should report that it found no relevant evidence. This is a useful test of whether the agent invents content to fill the brief.

Add live sources as a separate step.

Once the supplied-note example works, add collection before extraction. Configure the tools in your local Codex environment. Save the source URL, title, publication date, access time, and the text that supports the finding. Stop or record a gap when a source cannot be read.

For a weekly brief, pass an explicit date window. Decide whether a paper revision counts as new work. Prefer the paper, lab release, or original report as evidence. A search snippet alone is not enough to establish what the source says.

Keep each run and each change.

Inspect the source set, extracted findings, and final brief in the run records. If an unsupported claim passes, improve the extraction check and test the same input again in a new run directory. Keep the previous result for comparison.

Local runs and scheduling

The SDK executes on your computer and uses its available tools. This example has no built-in scheduler. Method Cloud can run methods on a schedule for pilot customers. Talk to Method about the source access and run schedule your task needs.