An agentic company

This is what we want Exosphere to be: an agentic company.

That means starting with primitives — the small, durable units of work that make a company go. Answering a support ticket. Reconciling an invoice. Triaging a bug report. Writing a release note. Most companies hire to do this work and think about automation later. We want to invert that order: design the work around agents first, and bring people in for the things only people can do.

Primitives, not jobs

A “job” is a shifting bundle of responsibilities — fuzzy, political, hard to measure. A primitive is sharper. Reply to this email. Decide which of these three customers to call. Summarise this week’s product changes. You can write a spec for a primitive. You can hand one to an agent and check the result.

So the team builds in two directions at once:

• Down to the primitive. Break each operational task into the smallest unit that’s still useful on its own.
• Up to the agent. Wire those primitives into agents that can run them end-to-end, with the right tools and guardrails.

It’s tedious. It’s also the only way we’ve found to keep the company from quietly accreting headcount as it grows.

Self-improving, not just smart

The hard part isn’t getting an agent to do something once. It’s getting it to do the same thing slightly better the second time, and meaningfully better the hundredth time. That’s where self-improving does real work:

• Agents log what they did and what happened next.
• They rewrite their own prompts and tool-use patterns when outcomes drift.
• They escalate when they’re not confident, and the escalations become the training data for the next iteration.

None of these mechanics are individually new. Wiring them together as the default operating layer of a company — instead of as one-off projects bolted onto an existing org chart — is the bet.

Side note

This is not a “no humans” pitch. It’s a “very few humans, all doing leveraged work” pitch. We think the right team for an agentic company is small, senior, and unusually opinionated about systems.

The agent does the primitive. The team designs the agent. The company is what comes out the other end.

What we’ll write about

This blog is partly a notebook for that bet. Expect:

• Notes on which primitives have been easiest and hardest to hand off.
• Honest write-ups when an agent loop gets stuck, and what we tried.
• The pieces of infrastructure we’ve had to build because nothing off-the-shelf quite fit.

If you’re thinking along the same lines — or think we’re wrong about it — we’d like to hear from you.