Inside the Refolk agent stack: plain-English briefs, live web crawling, and outreach grounded in real artifacts

Refolk presents as a single text box. You type a brief in plain English, the agent works for a few seconds, and you get a shortlist of people with drafted outreach for each. The interface is deliberately spare.

What is actually happening between the brief going in and the shortlist coming out is a small but specific stack: a query planner, a fan-out across live web sources, a ranking pass, and an outreach drafter. None of these are new on their own. The interesting part is how tightly they are coupled.

This is what that loop looks like from the outside, based on Refolk's public messaging and observable behavior in the demo on its site.

Step 1: parse the brief into structured constraints

The agent takes a brief like "senior Rust engineers in SF, ex-FAANG, now at small teams" and decomposes it into the constraints a structured search engine would need:

• Skill signal: senior, Rust.
• Geo signal: San Francisco / Bay Area.
• Career history signal: at least one stint at a FAANG-tier employer.
• Current employer signal: small team (which is itself a fuzzy constraint that has to be resolved against company headcount data).

This step is doable with a sufficiently capable language model and a careful prompt. The trick is making it deterministic enough that the same brief produces the same shape of search twice in a row, and making it loose enough that a poorly-worded brief still returns useful results.

Step 2: fan out across live web sources

This is where Refolk's bet differs from a vector-DB-over-resumes approach. Instead of pre-indexing profiles and serving from that index, the agent goes out to GitHub, LinkedIn and the open web at query time.

The advantage is freshness. A profile that was updated yesterday, or a repo that was first committed to last week, is visible immediately. The disadvantage is latency and cost. Each query has a real fan-out budget, which is part of why agentic sourcing tools tend to cap result counts and ask the operator to iterate the brief rather than scroll through thousands of results.

For technical roles, GitHub is the heavy source. Commit history, language breakdown, repo ownership, contribution graphs, conference talk transcripts, and the occasional README all become input. For non-technical roles, LinkedIn and the open web do more work.

Step 3: rank candidates against the brief

After fan-out, the agent has a candidate set that needs to be ordered. Two signals do most of the work:

1. Direct match against the brief's constraints. Skills mentioned, employers in history, geography.
2. Strength of supporting artifacts. A senior Rust engineer with 314 commits to a well-known crate ranks higher than one with five commits to a personal repo, all else equal.

The "line that earned them in" - the specific artifact that justifies a match - is a side effect of this step. The agent has to surface the artifact anyway to justify the ranking; once surfaced, it is reused as the seed for the outreach draft.

Step 4: draft personalized outreach

Outreach drafting is the part that has historically been hardest to automate well. Templated mail merge is easy and produces obvious spam. Genuinely personalized outreach requires reading the prospect's actual work, which is expensive for a human and only recently cheap for a language model.

Refolk's drafting prompt almost certainly does something like: given this prospect, the brief, the highest-signal artifact attached to them, and a short tone guide, write a four-to-six-sentence email that references the artifact specifically and proposes a single next step. The output reads like a careful sourcer wrote it, which is the entire point.

The model used is not disclosed, but the output quality and latency are consistent with a frontier-class model in the loop, fronted by a smaller model for the cheaper steps.

What this stack does not do

Two things worth being clear-eyed about:

• It is not a CRM. The agent finds and writes. It does not own the multi-touch sequence, the meeting booking, or the ATS handoff. Teams stitch those in with their existing tools.
• It is not a vector search over resumes. There is no "upload a JD and we will find similar resumes in our database" step, because there is no database in the traditional sense. The agent goes back to source on each query.

Both are deliberate. Stack discipline matters at this layer. The tools that try to be the entire sourcing-to-hire pipeline tend to be mediocre at all of it.

Why the architecture is the moat

Lots of teams can build a chat-with-resumes product. The reason agentic sourcing is harder than it looks is the fan-out and the grounding. Going to real web sources at query time, getting clean data back at acceptable latency, ranking on signal rather than keyword overlap, and surfacing the specific artifact that justifies the match - each of those is its own systems problem.

Refolk's bet is that the company that gets the loop tight wins the category, because the operator-facing UX is the same simple text box for everyone. The differentiation lives one layer down, in the quality of the fan-out and the grounding.

It is an architecture bet that looks correct so far. Whether it stays correct depends on how fast frontier-model costs fall, and on whether enough operators move their daily sourcing into this kind of loop to make the data flywheel real.

Track product launches, AI infrastructure news, and labor-market moves in real time on Quantisnow.