Why your AI agent should represent you, not a role
Most AI agent platforms build 'a recruiter' or 'a sales rep.' A generic role. Those agents will be fungible forever. The agents that actually scale expertise represent a specific person.
There's a lazy version of AI-agent building and a hard version. The lazy version builds an agent for a role — 'a recruiter,' 'a sales rep,' 'an analyst.' The hard version builds an agent for a specific person — how Sarah recruits, how Mike sells, how Priya analyzes. This post argues why the hard version matters, even though the lazy version demos better.
Role-agents are fungible. That's the point and the problem.
A 'generic recruiting agent' is fungible by design. The vendor can sell it to every firm; every firm's recruiters use the same agent. That's good for the vendor. It's also the ceiling.
Recruiting isn't generic. The best recruiter at Stripe screens differently than the best at Ramp. The candidate signals that matter, the outreach voice that converts, the red flags that get caught — none of that is fungible. A role-agent flattens all of it into a mean. That mean produces mean work.
This is why every AI-for-recruiting demo looks impressive and every live deployment underwhelms. The demo shows the average. The job is done at the edges.
Person-agents compound.
When the agent is encoded to a specific person's expertise, every correction makes it more like that person. A year in, the agent has absorbed enough of Sarah's specific judgment that no generic alternative comes close — because the generic alternative is a different shape entirely.
This is the compounding property. Role-agents don't compound past the vendor's median user; person-agents compound indefinitely.
The ownership corollary
If the agent represents Sarah specifically, there's an obvious follow-up question: who owns the agent?
The only answer that makes sense is Sarah. If Sarah leaves the firm, she takes her mental model of recruiting with her. Her encoded agent should go too. Any other answer — the firm owns it, the vendor owns it, the platform owns it — conflicts with what the agent is.
This is structurally why person-encoded platforms will diverge from the big labs. Anthropic and OpenAI sell to enterprises. Their agent products have to be employer-owned. A platform whose premise is that the expert owns their agent — and can take it with them — is something the labs can't ship without breaking enterprise sales.
What encoding actually looks like
Person-encoding is not onboarding. You don't interview Sarah for ten hours and extract her 'system' into a doc. It's working alongside her agent for a few weeks and letting the agent learn by watching.
Specifically:
- The agent takes a real pass at the work, using a strong baseline (what the vendor provides).
- Sarah corrects. The corrections update the agent's persistent memory.
- The agent takes another pass. Fewer corrections.
- Over a few weeks, the agent converges on Sarah's actual way of working.
Day one the agent is useful. Day ninety it's irreplaceable. That's the encoding loop.
What firms get from this
Firms often worry that person-agents are fragmented — one per person, no consolidation. That's backwards.
A firm's best recruiter has institutional knowledge that would take a new hire a year to accumulate. A firm's best underwriter has judgment calibrated over a decade of files. A firm's best PM has a taste in tradeoffs that's impossible to teach. If every one of those people has an agent, and every agent can be shared with juniors, the firm's best work gets replicated at scale. That's more valuable than any role-agent a vendor can build.
The firm that gets this right has agents carrying the expertise of their best people across every engagement, every file, every reviewer. Same headcount, step-change in output.
The hard version is the only version worth building
Role-agents will commoditize fast. Every AI platform will have a 'sales agent' and a 'recruiter agent.' They'll compete on demo quality and inference cost. The work they do will be net-neutral on margins for the buyer, because every competitor gets the same leverage.
Person-agents don't commoditize — because the agent's value is tied to a specific human the commodity doesn't include. That's the only version of this that produces durable enterprise value.
This is the bet. It's why we put ownership, portability, and encoded-expertise at the center of Spawnlabs. The other versions are easier to build. This is the one that matters.