Layer 3 · Fleet & Harnesses

A directed agent workforce — not one agent, an org.

One clever agent is a demo. A company is an org chart. We run a Conductor over a set of department directors, and we spin specialist agents up in whichever harness fits the job — from a Python SDK loop to an edge function to a visual workflow.

The org chart

Conductor + directors + specialists.

The fleet mirrors a real company: a CEO-agent coordinating directors for Dev, Marketing, Ops, Legal, Finance, R&D, and Sales — each with its own scope, skills, and memory — calling specialist sub-agents underneath.

Conductor

Reasons across the fleet, proposes the next action, routes to directors.

Directors

Department heads — Dev, Marketing, Sales, Ops, Legal, Finance, R&D — each with scoped context and the authority to delegate.

Specialists

Narrow standalone agents — a market-intelligence agent, outreach generators, scorers — spun up on demand.

Human in loop

The operator confirms. The fleet proposes; it doesn’t self-authorize.

The harness menu

Many ways to stand an agent up — one default.

There is no single right framework. We keep a catalog of harness patterns and pick by fit, with a default standardization — Pydantic AI + a brain-api (FastAPI) pattern + a LiteLLM proxy + Postgres — so most agents look the same under the hood.

Containers

SDK agents

Long-running, tool-calling agents in their own container — the workhorse for enrichment, scoring, and research.

Serverless

Edge functions

Stateless, event-driven logic close to the database — cheap, fast, and easy to govern.

Visual

Workflow automations

Drag-and-drop automations for integrations and scheduled glue — one isolated tenant per client.

DAG

Declarative pipelines

Multi-step YAML pipelines for our repeatable builds — market-intelligence reports, onboarding, landing pages.

Coding loop

Grind / Claude Code

Self-iterating coding agents on a VPS for heavy, long-horizon build work — flat-fee, off the critical meter.

Memory-native

Cognee + Neo4j

When memory itself is the use case — long-lived persona, evolving facts — the graph store is the harness.

The Karpathy Principle

We build an iOS, not a museum of frameworks.

‘Six months is an eternity in this space.’ Our standing rule: default to strong, well-supported methods and refine our own house method rather than chase every hyped framework. Every new harness gets a skeptical evaluation window before it earns a slot — and rejected ones are logged with the reason so we don’t re-evaluate them next quarter.

The skill catalog

A capability is a skill, and skills are portable.

A skill is a packaged capability — how to do one thing the company’s way. The catalog is large, versioned, and — crucially — portable: a skill we built for ourselves can be anonymized and re-tuned for a client’s brand and data without rebuilding it from scratch. That portability is what makes a comprehensive client build fast.