Why an "AI-agents-first" company can't and shouldn't exist

"AI-agents-first" has become the new "mobile-first" and "cloud-native": a phrase people put in pitch decks to sound inevitable. We don't think it describes a real kind of company. We think it describes a category mistake, and one that gets more expensive the longer you build on it. Here's why.

Agents are tools, not customers

A company exists to solve a problem for someone who has it. An agent is not someone. It has no problem, no budget, no patience, no taste, and no consequence when things go wrong. Treating agents as the primary user means you stop asking the only question that keeps a product honest: what is the human actually trying to get done?

Agents are a layer in the stack, the way a database or a queue is a layer. Nobody builds a "Postgres-first company" and expects that to be a strategy. The tool serves the workflow; the workflow serves the person. Flip that and you end up shipping features that demo well to other agents and help no one.

It can't exist: accountability has nowhere to land

Software already fails in boring, predictable ways. When it fails for a paying customer, there is a chain of responsibility: someone wrote it, someone deployed it, someone is on call, someone owns the outcome. An agent-first loop breaks that chain by inserting a component that is, by design, not accountable. It can act, but it cannot be fired, sued, or paged at 3am in any meaningful sense.

• Reliability: agents are probabilistic. A company whose core product is "the agent decides" has no stable contract with its customers, only a distribution of outcomes.
• Cost: every interaction burns tokens, retries and tool calls. At scale the unit economics look like a subsidy, not a business.
• Compliance: regulators and auditors ask who did what and why. "The agent did it" is not an answer that survives a review.
• Trust: trust is earned by being boring and correct. Agents are, by construction, neither on demand.

It shouldn't exist: it replaces judgment with throughput

The pitch is that agents remove humans from the loop. But the loop is where the judgment lives. Monitoring, incident response, code review, support, hiring, writing: the value is not in producing more of it, it is in deciding what matters and what doesn't. Agents are excellent at the first and useless at the second, and pretending otherwise doesn't automate expertise, it just dilutes it.

There is also a simpler reason. Companies are made of people who chose to build something together. An "agent-first" org is an org that has decided its own craft is the thing to automate away. That is not a strategy; it is a slow exit. The people who care about quality leave first, and what remains is a wrapper around a model with a logo on it.

What actually works

Agents belong where they are boring and bounded: summarising, drafting, triaging, searching, generating the first pass of something a human will check. The company that wins is not the one that puts agents first, it is the one that puts the customer's outcome first and uses agents to get there faster, with a human still owning the result.

• Keep a human in the loop wherever the cost of being wrong is high.
• Measure agents by the outcomes they enable, not by how much they do.
• Make the agent's work inspectable, reversible and attributable.
• Build the boring infrastructure (monitoring, rollback, audit logs) before you build the clever agent.

The PulseWatch Team