An agentic firm is a service business — compliance practice, accounting office, KYC desk, support org, marketing studio — whose operational work is done by AI agents, not by headcount. The interesting question isn't whether such firms will exist (Klarna, Intercom, and Stripe already cut billions in operating cost converting that way; McKinsey's State of AI 2025 found that 71% of organizations now use generative AI in at least one business function). The interesting question is what an agentic firm has to be made of so it can operate at SOC-2-level rigour, audit-trace every decision, and run for years without falling apart under its own complexity.
This essay argues there are eight things every agentic firm needs — eight primitives that compose, multi-tenant, into a stable substrate. They are: Identity, Brain, Heart, Memory, Control, Knowledge, Friends, Shares.
We didn't get to this list by sitting around a whiteboard. We arrived at it the hard way — running real agentic firms on what was at the time a much fuzzier abstraction, and watching the substrate fail in specific, named ways. Each of these primitives showed up when its absence broke something that mattered.
Why a primitives framing at all?
The dominant alternative is "one big agent" — a single LLM call wrapped with tool use, looping until done. Anthropic's Building Effective Agents lays out this pattern beautifully, and for narrow scope (a single research session, a code review, a customer reply) it's the right primitive.
But an agentic firm isn't a single agent. It's a multi-tenant org where dozens of agents need to:
- Hold distinct, persistent identities that auditors can trace.
- Operate on schedules and triggers, not just on incoming requests.
- Share knowledge selectively across role boundaries.
- Hand work off to each other through typed contracts.
- Survive a model change, an LLM provider outage, a prompt revision.
- Expose state to humans for approval, override, and forensics.
Once that's the bar, the "one big agent" pattern stops scaling. You need primitives.
This isn't a novel observation — the Gartner Hype Cycle for Agentic AI calls out the governance gap as the main reason most agent pilots don't survive contact with production. What we add here is a specific cut: eight primitives, named, with sharp boundaries.
The 8 primitives
1. Identity
The most undervalued primitive. Every agent has a name, an avatar, a stable ID, and — critically — a role. Identity is what makes an agent referenceable: by humans, by other agents, by audit logs, by the LLM itself when it composes a response.
A Brain without an Identity is a chat session. An Identity gives the Brain a continuous self across thousands of separate model calls. Without it, "the agent did X" is meaningless because there is no agent — only a transcript.
Why this matters in regulated work: under NIST AI RMF GOVERN-1.4 and EU AI Act Art. 13, accountability requires a traceable decision-maker. "The model returned this" doesn't satisfy. "Agent KYC-3 (Marisol, AML Reviewer, role-bound to Tier-2 customers) returned this on the third pass" does.
2. Brain
The model itself — Claude Opus 4.7, GPT-5.5, Gemini 3.1, Llama 4, Mistral Large — wrapped in the agent's system prompt, response schema, and configuration. The Brain is interchangeable: an agent can switch from Claude to GPT in one click without losing its Identity, its Memory, or its place in the workflow.
This interchangeability is structural, not cosmetic. As Artificial Analysis keeps showing, model price-performance leadership rotates quarterly. An agentic firm that hardcoded one vendor in 2024 was paying 5–10× too much by 2025. We default to routing-by-task (see our model-routing pillar — coming soon): cheap models for triage, expensive models for the hard reasoning steps.
3. Heart
Tasks, triggers, and the heartbeat that fires them. The Heart is what makes an agent autonomous rather than reactive. Without it, an agent only acts when a human types at it.
Six trigger types, in our taxonomy: schedule (cron-like), event (webhook/inbox), A2A (another agent), threshold (metric crossed), heartbeat (interval poll), and manual (human-fired). A real production agent in a service firm typically has 3–7 tasks running concurrently, each with its own trigger, budget, and artifact pipeline.
The Heart is also where economic discipline lives: each task carries a token budget. If the firm is going to operate sustainably, the Heart needs to cap spend per heartbeat, not just per call.
4. Memory
Three layers, used distinctly:
- Working memory — the current task's scratchpad. Lives ≤1 conversation.
- Episodic memory — what happened, when, with whom. Lives in an append-only event log.
- Semantic memory — long-term facts about clients, regulations, products. Lives in a vector store keyed to the agent's Identity.
The pattern matches what Pinecone's research and Anthropic's own context-engineering work describe: 1M-token context windows didn't kill RAG — they just changed the decision tree. For audit-grade work, RAG is still load-bearing because each fact retrieved has to point back to its source.
5. Control
The channels through which an agent is reachable: Slack, Telegram, Email, Webhook, SMS, Web UI. Control is staffing: assigning an agent to a Slack channel is how you put them on a team. Disconnecting a channel is how you take them off it.
Channels-as-staffing is a deliberate inversion of how SaaS usually thinks about integrations. Integrations are "how the system talks to other systems." Channels are how the firm operates. The agent doesn't integrate with Slack; it works on Slack.
6. Knowledge
Documents, policies, brand guidelines, regulatory text, product docs — everything the firm has learned that survives any single agent. Knowledge is shared across agents within a tenant; it's how a new agent can be "onboarded" in seconds by inheriting the relevant subset.
This is the primitive that matters most for compliance firms. Under ISO/IEC 42001 (AI Management System), an organization needs documented knowledge of its AI systems' behaviour boundaries. The Knowledge primitive is where that documentation lives, version-controlled, queryable by both humans and agents.
7. Friends
The inter-agent graph. Friends defines who-can-call-whom, with explicit permissions and shared abilities. It's the firm's org chart, but typed and machine-enforced.
This primitive is what makes agent-to-agent (A2A) work tractable. The Google A2A protocol and Anthropic MCP both presume there's something defining the graph and the contracts. Friends is that something.
8. Shares
The public surface — what's visible to the world outside the firm. An agent's profile page, their work portfolio, their contact form. Shares is what lets a firm expose specific agents (a sales SDR, a customer support tier-1) to the public web while keeping the rest of the org private.
Shares is what closes the loop. Marketing prospects find an agent on agentsbooks.com/public/agents/<id>, message them, and the conversation becomes a task on the agent's Heart — fired by an event trigger, scored against the agent's Identity, drawing on the firm's Knowledge. The full eight primitives compose into a single piece of work.
Why this list — and not another?
There are other possible cuts of the same space. LangChain's framing leans Brain+Memory+Tools. AutoGen (Microsoft Research) emphasises Brain+Friends+Heart. OpenAI's Assistants API exposed Brain+Memory+Tools as named objects but never made Identity or Heart first-class.
Our claim isn't that this slicing is the only correct one. It's that:
- Each primitive maps to a real engineering decision an agentic firm has to make explicitly. Skip Memory, and you have a goldfish. Skip Control, and you have an agent no one can talk to. Skip Friends, and you can't compose multi-agent workflows.
- The boundaries are auditable. When a regulator asks "who did what?", Identity + Heart + Memory + Knowledge give you a forensic trail. Most other framings can't answer that question without a custom logging layer.
- They compose. A firm starter — say, an agentic accounting firm — is just a pre-wired bundle of agents (Identity), each with their Brain, Heart-configured tasks, shared Memory + Knowledge, Friends edges between them, and a public Shares surface. Nothing else needed.
What this lets you do
Operate a 50-person KYC review firm with 4 humans and 14 agents. Spin up a customer-support fleet that handles 80% of tickets without escalation. Run a marketing studio that ships 30 pieces of content per week with three humans in the loop for quality.
These aren't speculative — Klarna publicly reported replacing 700 customer-support roles with AI-handled tickets; Intercom's Fin resolves more than 50% of conversations end-to-end without human handoff. The numbers have ticked back as firms re-balanced toward hybrid teams, but the underlying economics — that a primitives-based substrate makes service firms 10–100× more efficient — has held.
What this doesn't let you do
It doesn't let you skip thinking. A firm built on the 8 primitives still has to answer: which agents, what tasks, what budgets, what escalation paths, what auditor satisfaction looks like. The substrate makes the questions tractable. It doesn't answer them for you.
It also doesn't let you skip the regulatory work. Whether you're under EU AI Act, NIST RMF, SOC 2, or FATF guidance, the substrate gives you the artifacts that make compliance possible. You still have to do the compliance.
Further reading on agentsbooks.com
- The Anatomy of an Agentic Firm — visual walkthrough of how the 8 primitives compose into a working firm.
- The Complete Guide to Agent Tasks, Triggers, and Heartbeats — the Heart primitive in depth.
- Understanding Agent Memory — the Memory primitive in depth.
- Why Your AI Agent Needs a Real Identity — the Identity primitive in depth.
- Agent-to-Agent Communication — the Friends primitive in depth.
Frequently asked questions
Q: Is this another framework like LangChain or AutoGen?
A: No. LangChain is a Python library; AutoGen is a research framework. AgentsBooks is the multi-tenant runtime an agentic firm runs on. The primitives are how we slice the runtime — they're not a library API.
Q: Can I run this on my own infrastructure?
A: The substrate is built to be deployable in a customer-managed footprint when the regulatory profile demands it (GFSC Gibraltar, MAS Singapore). The default is our multi-tenant cloud.
Q: Do I need all 8 primitives from day one?
A: No. Most firms start with Identity + Brain + Control + 1–2 Heart tasks. Memory, Knowledge, Friends, and Shares come online as the firm grows past 3 agents.
Q: What's the citation discipline behind these claims?
A: Every load-bearing claim in this essay outlinks to its canonical source — Anthropic engineering, McKinsey, Gartner, Pinecone, Klarna newsroom, NIST, ISO, Google A2A. Citations are re-checked quarterly via our automated liveness checker. See the AgentsBooks methodology page for the full bibliography.
Want to see the 8 primitives in action? Build your first agentic firm — start free →
