Observations — 0003 // The Company Harness
This essay is a continuation of a previous one: Observations — 0002 // OODA loops and the Latency Game.
The year is 2027. You want to launch a new company. So, you take the obvious next steps:
→ Pick a name, buy a domain, and set up email
→ Jump on Stripe Atlas and incorporate
→ Open up a bank account with Mercury and a credit card with Ramp
Then, one final step:
→ You set up your company harness
What’s a company harness?
You can’t go out and buy one today. But it’s the biggest missing layer in the AI stack right now — and very soon, it will be as essential to launching a company as incorporating or opening a bank account.
*****
Have you ever tried to create a fully agentic workflow for your company?
In my last essay, I said that agents need 2 things:
- A place to run
- Context to run with
That was accurate, but oversimplified. Once you start peeling back the onion, you realize that agents are very needy indeed, and that leaves you with a wall of technical choices:
| Environment | Where does the agent run? In a cloud sandbox? On a server? A Mac Mini in the coat closet? |
| Model routing | What model(s) do you use to balance cost and effectiveness? |
| Tools | What tools does the agent need to do its job? What's the right trade-off between safety and effectiveness? |
| Triggers | What wakes the agent up to go and do its work? (Crons, webhooks, etc.) |
| Observability | How do we know what the agent is actually doing, and what it did? |
And potentially also:
| Hooks and safeguards | Which actions do we want to enforce with code, or block? |
| Orchestration | If multiple agents are working together, how do they coordinate and communicate? |
| State persistence | What happens when an agent fails mid-run? How does it pick up where it left off? |
| Cost monitoring | How do you know what the agent is spending on tokens? What caps do you want in place? |
If that seems like a lot — it is! That friction you’re feeling in setting up even simple autonomous agents isn’t just you. Each choice in these tables has trade-offs, and figuring it out is a genuinely hard problem.
Collectively, these decisions describe the agent harness: the technical infrastructure that makes an agent run.
If you’ve ever used Claude Code, Codex, OpenClaw, Hermes Agent, etc. you’ve experienced an agent harness.
The agent harness will never be a solved problem — there’s always more you can optimize. But there is a robust set of primitives and a growing ecosystem of tools. And platforms like Pi make it accessible to customize your own.
*****
So then why is it still so hard to run agents in your business?
Because the agent harness only solves half the problem. It tells you how to run an agent — but says nothing about how to make that agent effective at your company.
For that, you need a second layer:
| Context and knowledge | What does the agent need to know about your business — your products, customers, processes, pricing, strategy — to make good decisions? Where does that knowledge live? |
| Integrations and permissions | What systems does the agent need access to — your CRM, your ERP, your email? What permissions does it have? How are credentials managed? |
| Escalation logic & UI | How does the agent know when to get a human involved? And where does it interact with the human when it does? |
| Learning and memory | How does the agent learn from past mistakes and get better over time? What institutional knowledge does it need to retain? |
| Domain-specific evals | How do we know the agent did its job correctly — not just technically, but by the standards of your business? |
| Meta-decisionmaking | How much do you require humans to make all of the decisions above, vs. leaving it with the agents themselves? |
You can have your choice of great agent harnesses today; but the market has yet to deliver a winner of this second layer. Right now, setting up a new business agent is a bespoke integration project every time. You cobble together the context, wire up the permissions, define the escalation paths, teach the agent about your business from scratch. Per agent.
That’s why most companies try one or two agentic workflows and then stall out. The upfront cost of each new agent is just too high.
There needs to be a layer that does this work once, for the whole company — so that spinning up the next agent is as simple as giving it a responsibility and pointing it at the harness.
That layer is the company harness.
*****
But to build one, you have to answer a messy question: what even is a business?
Today: it’s an agglomeration of people, power, responsibilities, conversations, decisions, reputations, regulations, trust, plans, conventions, memories, roles, strategies, tactics, experiments, structured and unstructured data.
— It’s the COO cutting up some data in Excel, eyeballing a chart, making a judgment call.
— It’s a salesperson hitting up their old clients at a new job.
— It’s the CEO winning the sale at a loss, just to piss off their rival.
— It’s the General Counsel making a judgment call on how a regulation should be interpreted.
— It’s a software architect turning the company’s values around data protection into code.
Some of this translates to files and code right now. Your product catalog, your pricing rules, your org chart, your approval workflows — these are already structured enough for agents to act on.
Other things — the CEO’s grudge, the GC’s interpretive judgment, the salesperson’s relationships — are irreducibly human. And they should stay that way. (I’ll argue in a future essay that humans will be more central to the AI-era company… just massively levered up.)
The company harness doesn’t need to encode everything. But it does need to make agents aware of the full landscape — including which decisions require a human, who that human is, and enough context to route well.1
*****
While nobody has cracked the code yet on the company harness, it’s clearly a need that the market is calling for. And many companies are chasing the prize. Every business that wants to run agents effectively needs to bridge this gap — and right now, most are doing it ad hoc, one painful agent at a time.
In my next essay, I’ll survey the landscape — who’s building company harnesses, what approaches they’re taking, and where the biggest opportunities lie.
Footnotes
-
A distinction worth making: the company harness is not the same as what some people have been calling a company brain — usually presented as a wiki-style knowledge base that tries to capture everything in one place. I am very bearish on wiki-style company brains. Static knowledge stores always drift from reality, can only encode one ‘cut’ of the information, and become monstrously expensive to maintain. In equal and opposite measure, I am bullish on graphs as substrates for encoding companies… but we’ll have to save that for another post. ↩