Multi-Step Research Agent

An Agent that plans its own searches, reads across sources, checks claims against each other, and writes a cited report. With the evaluation set and cost budget that keep it honest.

The problem

Questions like "what changed in data-residency rules for our region this year, and what do we need to do?" take a day of reading and still end with an unsourced summary.

Who it's for

Analysts, consultants, operations leads who produce written recommendations.

What you get

A report of 600 to 1,200 words with a claims table: each claim, the sources that support it, the sources that contradict it, and a confidence note. Plus a run log of every search and read.

How the workflow runs

01 Question
02 Plan searches
03 Read and extract claims
04 Cross-check claims
05 Cited report

Tools and accounts

A workflow tool or small script with loops and a model step (this project uses a loop)
A web search API
A place to store the run log and the report

Before you start

Evaluating Agent output
Cost and latency budgets
Retries and error handling

Result first

Input: "What changed in data-residency rules for cloud storage in [region] this year, and what would a 50-person company need to do?"

Output, about 900 words, ending with:

Claims
| # | Claim | Supports | Contradicts | Confidence |
| 1 | Rule X took effect in March | [2][5] | - | high |
| 2 | Applies to companies over 25 staff | [2] | [4] says 50 | low: thresholds differ by source |
| 3 | Existing contracts have a 12-month transition | [3][5] | - | high |
Not resolved: the threshold in claim 2. Sources [2] and [4] are both official but dated differently; the newer one is [4].
Run: 6 searches, 9 reads, 41,200 tokens, 3 min 10 s.

The report says what it knows, what it does not, and what it cost to find out.