Build a Sourcing Agent That Learns From Your Feedback With the Wrangle API and MCP
Most sourcing tools return a list and stop. Wrangle's Agents already cover the recurring part: give one a standing brief and it reruns on a schedule and surfaces only the candidates you haven't seen. This post is about a different problem. How do you turn the decisions you've already made (every hire, offer, and pass) into criteria, and keep those criteria current as you review?
Wrangle's API and MCP server make that possible, and Serval built a loop like this around their own recruiting process. Here's the pattern and how to build your own version.
What you'll build
Four pieces:
- A criteria document built from your past interview decisions.
- Scheduled searches that call Wrangle with those criteria.
- A quick review step: a thumbs-up or thumbs-down, plus a one-line reason.
- A write-back step that updates the criteria from your verdicts.
Wrangle handles search and enrichment. Your own tooling handles the criteria, the review, and the write-back, which keeps the loop tied to your judgment instead of a generic idea of a good candidate.
Choose your connection: API or MCP
REST API. People search, plus email and LinkedIn enrichment, as plain REST endpoints. You get back scored, structured candidates. Authenticate with an API key. This is the better fit when your own system, like an automation platform or a script, kicks off the searches.
MCP server. A remote Model Context Protocol server with tools for candidate search, full profiles, collections, and connected sources. Connect it to ChatGPT, Claude, or any MCP-compatible client and authorize with OAuth. This is the better fit when you'd rather have an AI client run the loop conversationally.
Both work against the same workspace as the app, so whatever an agent finds shows up where your team already works. Usage is metered in credits and you can check your balance in the dashboard, so test with small batches first.
Step 1: Turn past decisions into criteria
Your interview history is the best brief you have. Gather the candidates you hired, the ones who got offers, and the ones you passed on, along with your notes on why.
Then ask an LLM, or do it yourself, to boil the patterns down into three lists:
- Must-haves: traits every hired candidate shared.
- Nice-to-haves: traits that showed up often but not always.
- Clear misses: the reasons candidates were passed on.
Add target companies based on vertical, product line, and infrastructure similarities. Keep the result as a plain document you own. Every other step reads and edits it.
Step 2: Run searches from the criteria
Send the criteria to the people search endpoint, or ask your MCP client to run the search. Wrangle scores every profile against the whole brief instead of matching keywords, so a detailed criteria document gives it far more to work with.
Pick your batch size up front. Serval's workflow pulls 50 candidates each morning and another 50 each evening. If you're new to this, start smaller and adjust as you learn. Serval's talent lead eventually slowed his scheduled search from every day to twice a week because candidates were arriving more often than he wanted.
Step 3: Review with reasons
Show each profile with one question: yes or no. Add a short reason to every no, like "too early stage" or "no infrastructure experience." Wrangle shows the searches it ran, the sources it read, and the logic behind each score, so you can usually see quickly why someone ranked where they did.
Send approved candidates to your ATS. Wrangle connects with Ashby, Greenhouse, Lever, and Recruit CRM, and through MCP you can also save candidates to a Wrangle collection.
Step 4: Write verdicts back into the criteria
This is the step that makes the loop learn. After each review session, update the criteria document from your verdicts:
- Keep rejecting one type of company? Add it to the clear misses or take it off the target list.
- One profile type keeps getting a yes? Move its traits up to must-haves and add similar companies.
- Same rejection reason three times? Make it an explicit criterion.
The next scheduled search then runs against the updated document. If you use a Wrangle Agent, update its standing brief with the new criteria. Either way, your decisions flow back into the search without you rewriting it from scratch.
Tips
- Start with one role. Get the loop working on a single search before you add more.
- Log the why. A one-line reason on each rejection is the most useful input the loop gets.
- Read the criteria weekly. The loop edits a document you can read, so check that it still says what you mean.
- Stay in the decision. The automation brings candidates and context. You decide who moves forward.
Get started
Wrangle has a free trial, and teams can be up and running in as little as five minutes. Create an account, generate an API key or connect the MCP server, and run one search against a criteria document built from your last few hires.
