
Harish Deivanayagam
•26 days ago
GTM work is increasingly done by agents: AI SDRs, recruiting agents, deal-sourcing copilots. The open question is what those agents run on.
The default answer today is a script over a snapshot. An agent scrapes a source, dumps rows into a table or a CSV, enriches some of them, and works off that file. It works once. Then the export ages, the emails bounce, and the parser breaks the day the source changes its markup.
The better substrate is a live graph: people and companies indexed continuously, resolved to one identity, searchable with plain intent, and re-verified while you work. That is what Monial Sets is. This post is the argument.
An agent acting on a person needs three things to be true.
The row is the person. One identity across LinkedIn, the company site, filings, and the rest of the public trail — not three half-profiles the agent reconciles with a guess every run.
The fields are current. A job title from last quarter produces an email that congratulates someone on the old role. A verified email is one that worked recently and gets checked again — not one scraped in 2024.
The data is structured. The agent should read JSON it can act on — role, company, tenure, contact data — not re-parse prose into a schema on every call.
A script over a stale table gives you none of the three reliably. It gives you whatever the snapshot was, at whatever resolution the scraper managed, with dedupe as a weekend project.
Four ways the substrate fails when it is homemade:
Sources change. A scraper is code written against someone else's markup. Every redesign is a silent breakage, and the person who notices is a rep with an empty list.
Snapshots age. The table you built Monday is not the market on Friday. People change jobs, companies raise, inboxes die. Nothing in a CSV tells you which rows stopped being true.
Identity is the hard part. "Is this the same person?" across five sources is the whole problem in people data, not a utility function. Every team that rolls its own resolution rediscovers that.
Every new question is a new script. "Who became a VP of Sales last month" and "which of our targets just filed a C-corp" are two scrapes and two parsers today. On a graph they are two queries against the same resolved entities.
The same motion over an API looks like this:
The agent asks for staff-level infra engineers at 200–1,000 person fintechs who changed jobs in the last 90 days. Sets returns resolved, structured rows — 0.5 credits a result. The agent enriches the 40 it will actually contact with verified personal emails — 5 credits a contact — and enrolls them in a sequence with per-account daily caps, sends staggered across working hours, and stop-on-reply at 0.5 credits an action. A watcher on each company pushes job changes and funding events to the agent's webhook the day they happen.
No scraper to maintain, no snapshot to age, no dedupe to improvise. The judgment stays in the agent; the substrate is metered infrastructure.
A script is the right tool for one-off research you will not run again, for bespoke parsing nobody else needs, and for prototyping a question before you pay to ask it at scale.
But keep the substrate external even then. Who exists, what is true about them now, and what just changed should come from a graph that is re-verified continuously — not from a file the agent rewrote last week.
The judgment: who matters, what the message says, when to act, when to stop.
The substrate: who exists, what is true about them now, what just changed. Monial's split maps to that — Sets is the substrate, Sequencer is the acting, Monitor is the knowing. Hand your agent the graph, and keep the scripts for the parts that are genuinely yours.
If you want to see how coding agents call this in practice: Claude Code for GTM in 2026. The honest comparison with Deepline, Crustdata, Coresignal, and the contact-data tools is Monial vs others.