Built from the practice we run inside the studio.
Workshops on weekends, on-site consulting on weekdays, retainers for teams who want a sparring partner past the kickoff. Three engagement shapes — agentic-CLI training, OpenClaw and Hermes Agents setup, and the vibe-coding / harness / context / memory coaching that gets a team past the demo.
Hands-on training on the three agentic CLIs your engineers will actually touch this year.
Install, identity and credentials, project setup, slash commands and skills, MCP servers, hooks and settings, plan vs. agent modes, sub-agents and worktrees — and the things that bite a team in week two: cache misses, runaway tool loops, context drift, prompt-injection in retrieved content.
Half-day intro workshop, two-day full workshop, or a four-week embedded coaching block where we sit with your engineers through real work — not a sandbox, not a kata. The codebase is yours; the tickets are real; the discipline is what we are teaching.
Tool-vendor cheerleading. We use these every day; we tell you where each is strong, where it's weak, and where to use which. No "AI will replace your team" deck — that talk is not useful in week two when the tool has just blown up a migration.
For teams standing up an internal agent platform instead of buying SaaS.
- When this fits. You've decided you need an internal agent platform — an action bus, a memory tier, an eval harness, a refusal policy, an audit trail — and you want to operate it yourself rather than buy a closed SaaS product on a per-seat curve.
- What we bring. Reference architectures we've shipped, deployment templates, the boring decisions — auth, secrets, networking, observability, audit — settled before the first crew runs. We've made the mistakes already; you skip them.
- How the engagement runs. A four-to-eight-week setup engagement, paired with an optional retainer for the first quarter of operation. We hand it back to your team — we don't squat. The platform is yours; the runbook is yours; the on-call is yours.
- What it does not include. We don't bundle a managed-service contract on top of someone else's open source. If a vendor relationship makes more sense than self-hosting, we will tell you, and that's a different conversation.
The four disciplines that separate a working agent from a clever-looking one.
- AI vibe-coding discipline. When to let the agent run; when to step in; how to read agent output without rubber-stamping. The daily-cadence habits that decide whether the team gets compounding leverage or compounding mess.
- Eval-harness engineering. Golden, aspirational and adversarial sets. Writing evals that catch regressions instead of celebrating happy paths. Costing the harness and budgeting it like a test suite, not a vibe.
- Context engineering. What to load, when to compact, when to clear. Cache-aware prompt structure. Retrieval that earns its tokens. The unglamorous work that decides whether your agent is fast or slow, cheap or ruinous.
- Agent memory tiers. Working memory, session memory, long-term memory, organisational memory. Where each lives, what each costs, how each decays — and which decisions belong to a human review, not a memory write.
- Who joins. Engineering manager, not just engineers. The discipline doesn't stick without manager buy-in — the new failure modes are managerial as often as technical.
Workshop bookings, embedded-coaching slots and platform-setup engagements are scoped during the paid diagnostic week. You walk away with a written scope and a calendar, regardless.
Brief us →