A sourcing agent in 30 minutes

Two years ago, “building a sourcing agent” meant a project: a developer, scrapers, a deduplication pipeline, weeks of tweaking. Today, the hard part — data aggregation and freshness — can be delegated to an existing index, and the agent part takes place in one work session. Here is the complete process, minute by minute, without writing a line of code. The only prerequisite: an MCP client (Claude Desktop is sufficient) and access to a recruitment data index.

Minutes 0 to 5: plug in the index

An MCP server is declared in the client configuration: a URL, an authentication key. That's all. From there, the assistant has tools that it can call up on its own: search for companies, filter profiles, list active job offers, consult a school's training courses.

The important point is not the manipulation, trivial, but what it replaces: you export nothing, you import nothing, you do not build any local database. The index stays where it lives, continuously updated; the agent consults it on request. The data you use at 9 a.m. is that of 9 a.m., not that of the last export.

Minutes 5 to 15: define the persona, in writing

This is the step that everyone botches and which determines everything else. A useful sourcing persona is not a job title, it is a set of prioritized criteria. For example:

• Target : data engineer, 3 to 6 years of experience, fluent French.
• Positive signals : experience in a fast-growing company, school with an identified data course, experience in a sector close to yours.
• Exclusion signals : in position for less than a year, profile already contacted by your team.
• Context : your company, the position, the actual seniority range of the team.

Write this in natural language, in a document that you will paste at the start of the conversation. Ten minutes of writing here saves hours of sorting later: the agent will apply these criteria to each request, systematically — something no human does at 6 p.m. on a Thursday.

Minutes 15 to 25: research in dialogue

Run the first query: “Build a first list corresponding to the persona. » The agent then makes several calls alone: ​​he identifies the relevant companies (for example tech companies with 50 to 500 employees having recently published data offers - a signal that a team exists and is recruiting), cross-references with the profiles in the index, filters according to your exclusions.

The difference with a classic Boolean search is in the dialogue. You can answer: “Too many Parisian profiles, expanded to Nantes and Bordeaux”, or “Remove consulting companies”, and the agent reformulates his requests accordingly. Each iteration takes thirty seconds, compared to a complete search reconstruction in a traditional tool. In practice, three to four iterations are enough to go from a raw list of 200 profiles to a preselection of 25.

Minutes 25 to 30: the commented shortlist

Final instruction: “For the 10 best profiles, justify in two sentences the suitability with the persona and point out the points to check during the interview. » The result is not a list of names: it is an argued shortlist - why this profile, what signal brought it up, what gray area remains. This is precisely the deliverable that a senior talent searcher produces in half a day, and it is also a document that you can challenge line by line, because each statement refers to a piece of data in the index.

What these 30 minutes reveal

Notice what disappeared from the exercise: the sourcing application itself. No filter grid to learn, no proprietary interface. There remains a model, a protocol, and an index. All the value has shifted to the quality of the latter: its coverage, its freshness, its deduplication. A brilliant agent plugged into stale data produces stale shortlists — confidently.

This is the role that sourcing.sh has chosen: not just another application, but the index — around 123,000 profiles, 200,000 companies, 98,000 schools, 1.4 million job offers — aggregated via partner sources, continuously refreshed by agents, and exposed in MCP, in API or directly in your ATS. Flat rate: your 30 minutes as an agent do not trigger a credit counter. Agents now write to each other within half an hour; the index remains a profession.