We don't sell "an AI agent." We field crews.

A single agent doing five jobs is five prompts braided together — when it fails, you don't know which of the five failed. A crew separates the jobs: one narrow agent per task, one lead to route, one curator to tend memory, one human to approve. Each piece has its own eval, its own budget, and its own failure mode.

Generalist agents demo well and ship badly. The moment you put one in production with real telemetry the failure modes interleave: bad retrieval poisons drafting, bad drafting fools the checker, and the trace is unreadable. Crews stay debuggable because each agent has one job and one eval. We use the big models — we just don't ask any single instance to do five things at once.

Most "agent memory" is one context window stuffed until it overflows. We separate by lifetime: tier-01 (charter, pinned), tier-02 (long-term project facts, curator-written), tier-03 (per-task scratchpad, disposable), tier-04 (shared knowledge base, read by every agent in the crew). Each tier has its own retention rule, write policy, and size budget. Nothing leaks across boundaries without the curator's approval.

Memory rots if nobody tends it. The curator is a dedicated agent whose only job is to score what gets written, promote what proves useful, demote what goes stale, and prune the rest. Without one, long-term memory becomes a haunted attic of contradictory facts. With one, it stays small, recent, and relevant — which is the only state in which it earns its tokens.

Tokens aren't free. Latency isn't free. A 60-token prose handoff and a 22-token typed handoff can carry the same payload — but one parses on the wire and the other has to be re-read. We keep reasoning steps in natural language (where they belong) and push telemetry, status, and routing onto a typed channel. A crew that runs all day pays for every word it doesn't need to say.

The human is the final gate — approve, edit, or reject — not a step on every turn. Agents do the research, drafting, checking, and routing autonomously. The human sees a packaged outcome with citations and a diff, then signs off (or sends it back). For the Article Crew that's an editor on every piece. For BD Crew it's a person approving every send. For WaterDoctor it's senior wGrow engineers on the ground.

That's our comfort zone. We work with audited workflows, regulated data, and long-running loops — designed with PDPA, MAS, and IMDA in mind from the first whiteboard session. Every crew run is logged, every memory write is attributable, every human approval is recorded. The audit trail isn't a feature we bolt on; it's the substrate the crew runs on.

Model-agnostic by default. We build on Anthropic Claude, OpenAI, and Gemini, and pick per agent based on what each role needs — long-context reasoning, structured output, cost. Memory lives in Postgres and a vector store we control, not inside the model provider. If a provider changes pricing, deprecates a model, or goes down, we swap. No single vendor owns your crew.

Most engagements start in 2–3 weeks. We open with a one-week diagnostic — crew shape, eval plan, weekly cadence on the table at the end of it. From there the first crew is usually running against real data inside a month, behind a human gate. We don't ship single-agent demos, and we don't ship without an eval harness on commit.

Both. For most engagements we design the crew, ship it, and hand over with an eval plan you can run yourself. For deeper work — like our WaterDoctor engagement — two senior wGrow engineers embed with the client team on the ground: sensor data, control loops, regulator-readable reports. The crew model is the same; the operating model is different.