Agents

A network of specialist AI agents, each scoped to a function.

A network of specialist AI agents, each carrying the context that a function inside the company would normally hold. The engineering agent operates against the codebases and the depth of blitz-docs; marketing, design, and operations agents are in development. The substrate today is curated repos and prompts swapped into Claude conversations as the work demands. Custom Claude agents and direct API integrations are planned.

The engineering agent is the one that exists in structured form. Here is what that means in practice.

The engineering agent is grounded in the Blitz codebases — the frontend, the backend, the legacy library — and in the depth of blitz-docs. The docs are not a brochure; they document the data models, the order lifecycle, the integration surfaces, the formula AST, the working-time utilities, the notification channels, the announcement system. When the agent reads them, it inherits the operating context that a senior engineer would carry.

That depth changes what the agent does. Instead of generating boilerplate at the margins, it plans features, evaluates trade-offs, anticipates the side effects on the order state machine, names the right modules to touch, and writes code that respects existing conventions. It works less like a code generator and more like a colleague who has read the room.

The substrate is intentionally minimal: curated repos, well-shaped prompts, Claude conversations. There is no custom infrastructure, no agent framework, no orchestration layer. The depth is in the context, not the tooling.

Three more specialist agents are in development.

The marketing agent is grounded in the brand voice doc, the marketing strategy, the four-layer playbooks, and the page-by-page website specs. It carries the same posture as the rest of marketing — restraint, no needy language, the four veto tests applied to every word.

The design agent is grounded in past visuals, brand principles, and the design specs that accompany every page. Its job is to keep the visual language consistent as the site grows.

The operations agent is grounded in the admin handbook, the playbooks, and the configuration surfaces. It handles the routine work the platform doesn’t fully automate — application reviews, statement approvals, exception handling.

Beyond the four named specialists, the architecture is the substrate. Custom Claude agents with their own context files. Direct API integrations where the agents need to act, not just advise. Agent-to-agent calls so that a marketing change can ask the design agent for a visual treatment, or a feature plan can ask the engineering agent for an implementation. None of that is built yet. The order in which it gets built is reactive — whichever specialist needs structured form next, gets it.

The point of building this is not to replace hires that don’t exist yet, or to encode institutional memory for some future scaling event. The point is the shape of the team itself.

Blitz is a small operation. The aim is for it to produce the output of a much larger one, with every function connected cleanly. The agents are how that arithmetic balances. Each one carries the depth of a department; each is reachable; each one’s context evolves as the company learns. The team stays small. The output doesn’t.

A Commandos team — disproportionate impact through specialization and coherence, not headcount.