Building Production AI Agents with OpenClaw
What we learned putting an open-source, local-first agent runtime in front of real client systems.
This was an internal engineering session, not a public broadcast, and it was not recorded. What follows is the write-up — the outline we worked through and what we took away from it.
What the session covered
- What OpenClaw is, what it is not, and who actually maintains it today
- Why a local-first agent runtime changes the conversation with security-conscious clients
- How agents reach people — messaging surfaces instead of yet another dashboard
- The permission model: declaring exactly what a skill may touch, and why that matters
- Where agent projects fail after the demo, and the guardrails that prevent it
- How we decide between OpenClaw, a hosted agent platform, or a plain integration
How it ran
- 1
What OpenClaw actually is
An open-source agent runtime under the MIT licence — not a product from Anthropic, not a product of ours. It runs on hardware you control and talks to whichever model you point it at, which is the property everything else in the session depends on.
- 2
Who maintains it now
Its creator moved to OpenAI in late March 2026 and handed stewardship to a technical steering committee. We covered what that means in practice when you are recommending a dependency to a client: how to read the release cadence, and what to check before you build on it.
- 3
Local-first, and why clients care
Plenty of prospects cannot send customer records to a third-party cloud — not because of preference, but because of the contract they signed. An agent that runs inside their own environment turns a hard no into a conversation.
- 4
Skills, and the permission question
Agent capability is packaged as skills. The 2026 releases moved towards signed manifests that declare which filesystem paths, network endpoints and commands a skill may use. We walked through why an agent with unbounded shell access is not something to put near production.
- 5
Reaching people where they already are
The agents that get used are the ones reachable from the tools a team already has open. We looked at the messaging-surface approach and how it compares with building another internal dashboard nobody logs into.
- 6
Where it breaks in production
Demos succeed because someone is watching. We went through the failure modes we have actually hit — silent tool errors, missing approval steps, no audit trail, and cost drift — and what we now build in from day one.
Presented by
AI Engineering Lead
OpenMalo Technologies
Solutions Architect
OpenMalo Technologies
What we took away
- A local-first runtime is a commercial argument, not just a technical one — it unblocks clients whose contracts forbid third-party processing.
- Treat agent permissions the way you treat database credentials: explicit, scoped, and reviewable.
- An agent nobody can reach from Slack or WhatsApp gets the same adoption as an internal dashboard nobody opens.
- Governance of an open-source dependency is part of the technical decision, not a footnote to it.
- Most agent failures we have seen are operational — no approvals, no audit trail, no cost ceiling — rather than model failures.
Read more on this
Articles from our team on the same subject.
OpenClaw is an independent open-source project released under the MIT licence, created by Peter Steinberger and now maintained by its technical steering committee. OpenMalo is not affiliated with the project — we use it to build agents for our clients.
