HR agents, PM agents, coding agents, IT agents, finance agents, sales agents, procurement agents, legal and support agents. Eight domains sounds like eight projects, eight vendors, eight integration hells. It isn't. They share one core: and the core is the reason the count doesn't compound.
What all eight actually have in common
Different as the work is, every agent does the same four things. It takes an input from a system, decides an action, executes it, and records what happened. The differences are the tools and the guardrails, not the skeleton.
- Telemetry - the same ride-data patterns under a procurement agent and an HR agent. What was requested, when, and what it cost.
- The ledger - every action lands in an audit trail. Finance needs it; so does IT; so does anyone who has to answer a question six months later.
- Orchestration - multi-agent pipelines with human checkpoints. A coding agent hands a half-done PR to a human, not to another model.
The adjacent improvement is the compounding one. An IT agent doesn't replace the finance agent: it feeds it.
Why domains shouldn't be silos
The classic failure pattern is one pilot per department, each bought separately and never integrated. The ticket tells you the invoice exists; the invoice tells you the PO exists; but no system holds the thread. When all three run on one core, the thread is the core.
This is also why we've shipped multi-tenant SaaS and a four-application ops platform before the agents. Tenant isolation that cannot leak, financial correctness by construction, concurrency handled at the database level: those standards are the skeleton the agents sit on.
What this means for a client
If every agent were its own system, the eighth agent would cost more than the first, and governance would be eight control surfaces. On one core, the eighth agent reuses the harness: one place to deploy, version, observe and approve. The domains are a catalogue, not a collection.
The practical rule we use: a starting point, not a ceiling. We scope the first agent to your workflow. We ship it, measure it, and let the next agent prove itself against the same ledger. The stack compounds: which is exactly what an AI layer is supposed to do.