A sourcing plan is where most hiring strategies quietly die. The market brief says "this role is scarce"; the recruiter still opens LinkedIn and types the job title. This agent closes that gap — it converts market intelligence into the actual working artifacts of a search: an ideal candidate profile, paste-ready query strings, a ranked poaching list, and timing triggers.
I ran it against the same brutal role as the Talent Intelligence Agent case study: a Machine Learning Scientist in generative audio at an AI-native company in Cambridge, MA — a talent pool that numbers in the low hundreds globally.
- Must-haves: deep-learning research applied to audio/waveforms — not just NLP or CV; hands-on training of generative models (diffusion, autoregressive audio, neural codecs); production scale.
- Strong signals: publications at ISMIR, ICASSP, NeurIPS; prior work on generative music, TTS, source separation, or neural codecs; open-source audio models; music literacy.
- Adjacent pools: speech/ASR researchers ready to move recognition → generation; diffusion researchers from image/video willing to cross to audio.
- Calibration: target 4–10 years post-degree (most movable); PhD ~60% of the pool — but don't over-filter exceptional self-taught candidates with a strong model portfolio.
Four more tiers behind it: neural-codec specialists, speech/ASR crossover, generative image-to-audio crossover, and a PhD/publication academic pipeline — each with the geography and account-type notes to actually run it.
- Competitor litigation and funding stress shaking researchers loose
- Big Tech speech-team reorgs and shelved generative-audio efforts
- Vesting cliffs — researchers ~4 years into a grant at Amazon, Apple, DeepMind
- Academic cycles — spring PhD graduations; the weeks after ISMIR / ICASSP / NeurIPS acceptances
It also set a coverage expectation up front — roughly 30–60 direct-fit profiles in the Boston metro, a 50–80-name working list with adjacent tiers — and named what the search would miss: stealth profiles and people who don't list audio keywords, with a plan to close the gap through a local AI-music meetup and citation-graph mining of known researchers' co-authors.
Instead of stopping at a plan, the agent ran the queries against live LinkedIn and graded its own strategy — including the unglamorous mechanics most sourcing docs skip: parenthesized Booleans don't work on a basic account, result counts are a paginated floor rather than a census, and one geography code silently returns zero.
| Query | Volume | Precision |
|---|---|---|
| "music generation" | ~70–80 profiles | low — mostly noise |
| audio AND diffusion | ~90–100 | low–medium |
| "generative audio" | ~15 | medium |
| "music information retrieval" | ~50 | high — page 1 dense with real fits |
| "neural audio" | ~15 | highest — nearly all genuine |
The high-precision keys were narrow technical phrases no job description would suggest. The broad first pass undercounted the market — and the agent said so, revising its estimate upward and labeling the new number a floor, not a census.
Final read: ~12–15 research-grade direct fits in Greater Boston, of whom 8–12 are senior and recruitable — after subtracting the people it discovered were already at the target company. Its verdict: "a named-account search, not a funnel," with relocation likely needed to fill at senior level with real choice.
Most sourcing plans are static documents that flatter the strategy that produced them. This one ships with its own evidence: every estimate is tested against a live search, every gap is flagged instead of guessed, and the recruiter inherits not just a list but the method — which queries to trust, which to skip, and where the invisible candidates hide.
Want the full walkthrough? Email me — I'll run it on your hardest role.