What you'll be able to do after this guide: describe, step by step, what a sourcing agent does in one run, and explain why the steps come in this order.
A sourcing run has a fixed budget of time and tool calls. The most expensive thing it does is research. So the run filters out unsuitable candidates as cheaply as possible before researching anything, and it writes to the CRM only once it knows the result. Setting the status is the very last write, because status is the signal other agents act on.
Step 1: Pick candidates. Search on what the business is (industry, size, revenue, niche), as defined in your segment. Ask for more candidates than you need, because many will be filtered out next.
Step 2: De-duplicate, all at once. Before researching anyone, check every candidate against the CRM in one pass:
Search by website domain first. It's the most reliable match.
Then search by company name, to catch subsidiaries and rebrands.
If a match exists, check its status. Customer and Rejected mean stop.
Check the exclusion list.
If the record already has a research date, it's done. Skip it.
Only candidates that survive count towards the daily target. If too few survive, page further through the search results or widen the search and repeat. We learned to do this as a single pass after a run in which the whole first batch of candidates had been researched before. Search tools have no memory of what your CRM already holds and will return the same top results month after month.
Step 3: Verify the entity. Find the registered name and company number from the site's legal pages, then check the official register. If the company isn't active, score it 1, mark it Skipped, note why, and stop.
Step 4: Gather evidence. Revenue (from filed accounts where possible), current tools, accreditations, and the specific buying signal. Tag every finding with its evidence tier.
Step 5: Find the contact. Build a shortlist in your role priority order, then choose one primary contact: the person most directly tied to the buying signal. Check their preferred first name (an email address starting "dan." means you greet Dan, not Daniel). Record an email only if it's confirmed or published. Find profile links through ordinary web search; don't browse social platforms automatically.
Step 6: Score the fit. Apply the rubric, which needs the contact details from step 5. Scores 1 to 3 are recorded and marked Skipped, with the research date set because research is complete. If the prospect type is Unknown, leave the company at Researching for a person. Scores 4 and 5 continue.
Step 7: Draft the message. Browse the prospect's own website, match what you find to a specific value angle, and draft the first message. If nothing specific turns up, say so in the note and use your general positioning. Never present a generic angle as tailored.
Step 8: Write to the CRM. In this order: company record, company fields, person record, then one note per topic. Check for an existing sourcing note before creating one.
Step 9: Set status last and hand off. Write the routing note with the full draft, then set the status to Review. Review means every field and note is in place.
Step 10: Log and summarise. Append the run's counts to the funnel log, write a one-line human-readable entry, and reply with a summary (see Logging and Run Summaries).
The agent's final reply is how a person checks its work. Ask for:
How many candidates were searched, how many were removed at each filter, how many researched, how many reached Review.
For each new prospect: the score, the key signal and its tier, the contact, the routing decision and why.
Whether the daily minimum was met, and if not, what ran out.
Anything unusual: a tool that was down and which companies it affected, a field it couldn't write, a candidate near an exclusion line.
Before reporting finished, the agent checks that every candidate it touched is in at least one of these states:
Research date set (complete, whether Skipped or Review).
A routing note written and status Review.
Left at Researching with the reason named in the summary (for example, prospect type Unknown).
An explicit outage flag in the summary, with the research date left blank so the backfill agent retries it.
A candidate in none of these states has fallen through a gap. That check catches gaps before they become invisible.
Next: Write the Spec and the Task Prompt.
