An AI teammate that actually learns your business.

A chat session forgets you the moment you close the tab. Launchpad Agent runs on a schedule, reads named sources, writes to named outputs, has a memory layer that survives sessions, and reports back on a canary field every run. You don't talk to it — you delegate to it.

The other difference is the architecture. A single Claude tab is a generalist. Launchpad Agent is a small team — an orchestrator that decides what to do, plus specialist sub-agents that each do one thing well. That separation is what stops the "drifted off-task" failure mode.

By default, a dedicated UK VPS we provision and manage. The agent is a Hermes-based deployment with our own opinion-set layered on top (the five-part job description, the orchestrator pattern, the reliability gate). The model is whichever current-generation frontier model best fits the workflow — usually Claude Opus or Sonnet for language-heavy work, GPT-class for structured planning. We treat models as swappable; you don't get locked in.

If you prefer, we'll deploy inside your own Microsoft 365, Google Workspace, or Azure tenant. The architecture is the same; the hosting boundary moves.

It can't replace work that requires physical presence, regulated professional judgement, or commercial negotiation. It can prep the email; it shouldn't sign the contract. It can draft the quote; it shouldn't issue it without sign-off on anything over your threshold.

We write the "won't do" line into the job description before the build starts. If a workflow doesn't have a clean "won't do" boundary, that's a sign the workflow isn't ready for an agent — and we'll tell you that before quoting.

A one-hour review. We go through every agent against its job description and its success condition. Hours recovered, canary stability, exceptions raised, corrections you made, weekly spot-check results. If the agent isn't earning its monthly fee in time saved, we tell you — and either rewrite the job description, retire the workflow, or refund the last month.

We don't keep clients on retainers that aren't paying back. It's how the model survives.

Yes. Every agent runs against a memory layer that includes your terminology, your preferred phrasing, your past corrections, and your decisions log. The first two weeks of any build is heavy on this — we sit alongside the people doing the work and capture the tacit knowledge that lives in their heads.

If your business has a substantial body of internal knowledge (documents, transcripts, decision logs) that needs to be made queryable, that's a KnowledgeForge conversation — the context-engineering retainer that sits underneath the agent layer.

Zapier and Make are rule engines — "if X happens, do Y." They're brilliant for deterministic workflows where the answer is always the same shape. Launchpad Agent is for judgement work — "is this a real enquiry or a duplicate? Which draft response best matches our voice? Should this quote use the standard template or the premium one?"

In practice the two work together. The agent makes the judgement call; Zapier or Make moves the result to the right system. We'll often deploy both.

That's the recommended path. Start with one agent on the workflow that's costing you most time. Run it for 90 days through the reliability gate. If it earned its keep, add a second — and you're into Team tier without redoing the foundation. The architecture is the same at every tier; the agent count, runs and on-call cover are what change.

You own the job descriptions, the memory layer, the eval dataset, and the agent definitions. We hand them over in a portable format (Markdown for the JDs and memory, JSON for the evals, plain Python or TypeScript for the agent definitions). The hosting moves with you, or we run it out for 90 days while you find an alternative. No lock-in, no exit fee.