08 / Problem Statement Internal
  • internal
  • cli

AI agents can do enormous work — but without the structure of a real company around them, they just freelance.

A company brain for every venture.

PROJECT
Lodestar
STARTED
2026
AUDIENCE
Me, for now
01 / The Problem

An AI agent without context is a contractor with no manager, no team, and no filing cabinet. It can do the task in front of it, but it forgets, freelances, and can't be held to anything. The work doesn't compound from week to week.

What a venture actually needs is the structure of a well-run business: departments with specific jobs, a place where decisions get filed, a review loop where someone notices what's stuck. Real companies are productive because that structure exists around the people. AI agents need the same shape — just built for one operator instead of a hundred employees.

02 / The Solution

Lodestar turns any venture's repo into a company brain. Eight specialist agents stand in for the departments a real business would have — chief-of-staff, customer, marketer, seller, engineer, designer, analyst, operator — backed by a state directory that holds decisions, events, and context the agents can actually reason over. Slash commands handle the weekly mechanics. A local dashboard is the executive surface. Every venture gets the same shape, run on a single Claude Code subscription instead of per-call API spend.

Eight departments per venture
Chief-of-staff, customer, marketer, seller, engineer, designer, analyst, operator — every venture gets the same org chart, so context and review move across ventures without re-orientation.
State the agents can reason over
Markdown for decisions and notes, JSONL for events, SQLite as the searchable index. The institutional memory of a real company, in a directory.
Runs on Claude Code, not API budget
Every agent invocation routes through the operator's Claude Code subscription. No per-call pricing, no separate billing surface to defend.
An executive surface
A local dashboard for what's moving, what's stuck, and what's been decided — the view the operator actually wants when they sit down on Friday.
Slash commands for the weekly mechanics
Reviews, decision logs, approvals — the recurring work that erodes focus gets reduced to a verb.
03 / The Outcome

Each venture runs like a well-structured small company. Its departments know what they're doing this week, the decision log answers what was tried last quarter, and the operator's job moves up a level — from doing the work to steering it. A Friday review across the whole portfolio takes under thirty minutes and ends with a clear bet on where to lean next.

04 / What We Learned
  1. 01An AI assistant alone isn't a business. Without a place to file decisions, a review loop, and a shape that holds across weeks, even a capable model freelances itself into incoherence.
  2. 02Meta-tooling earns its build cost only if the portfolio gets more coherent because of it — not if it's a nicer surface for what already happens in scattered files.
  3. 03An audience of one is a real product constraint. The point isn't that it generalizes — it's that it can be sharp for the specific operator running the portfolio.