"Seven days with OpenClaw. On the eighth, I forged a new engine."
FrontispieceThe transition from a tool (Solve this task) to a partner (Create this outcome).
Observation: In 2026, the engine ceases to be a hand-cranked apparatus and becomes a self-aware agent.
Precise input. Predictable output.
Given the destination, it arrives at an outcome.
The Shift → The Marvels → The Failures → The Rebuild → The Future
The workspace exposes the machine's instructions, memory, and daily trace.
Inspectable. Versioned. Durable.
The engine remembers in plain sight.
You steer every valve by hand.
State the outcome. The machine arrives at that outcome through trying, observing and trying again.
Automatic snapshots. Context preserved before compaction.
Long-term knowledge loaded into each main session.
Name, description, and gates define the fitting.
Capability arrives as text, not code changes.
Write doctrine. Slot it in. The engine gains a new trick.
"This is the future."
"It actually reasons."
"The openclaw is brilliant."
I was sold.
When a collaborator fails, it hurts more.
This is not a tool crash.
It is trust venting into steam.
[Loop Detected]
OpenClaw: *reloads*
OpenClaw: *restarts*
[Status: Stall]
One bad edit → restart loop → context lost.
Fragile config destabilises the entire engine.
Plate XIIAgent, browser, files, and credentials sit in one blast radius.
Insecure deployments, exposed gateways, attack traffic in the wild
Too much power in one boiler.
Plate XVKeep the paradigm. Rebuild the chassis.
Plate XVIIIIngredients: Elixir, Phoenix, Firecracker, PostgreSQL
Isolated environment. No direct access to secrets or workspaces.
Per-user encrypted vault. Agent never sees credentials.
Per-user workspace container. No cross-container communication.
Secrets never enter here. All external access via MCP only.
OTP supervision restarts only the broken part, leaving everything else running.
The agent inside is Elixir. It handles all workspace interactions.
Same principle: restart only the smallest affected part.
No full-page reloads. No full system restarts.
Granular fault recovery. The rest keeps running.
Plate XXIUser explicitly unlocks. User explicitly locks.
Locked = no long-running process can use a secret.
Go to sleep. Nothing breaks.
Push notification to phone → unlock for task or time span.
Your secrets. Your control. Even when you're away.
Plate XXIVEverything is local. Agent has full access to host, browser, passwords — everything exposed to the agent.
One breach, full exposure.
We use the browser as our battle-tested security boundary.
Agent runs in a hosted VM. We handle updates, firewalling, secure storage, secret management.
Throw desired outcomes over the fence. We do the rest.
Plate XXVErlang was built for telephone switches — millions of concurrent connections, zero downtime.
Today we plug users and agents into LLMs. Same problems. Elixir is the answer.
Solving engineering problems with non-deterministic engines (agents) is as hard as herding cats (developers) — just faster.
Amazing fit. Processes, supervision, fault tolerance — it just works.
Elixir works extremely well with AI coding environments.
Fast, highly dynamic frontends without the JavaScript complexity.
Plate XXV
"Mind the steam."
getahead.now
Scan to join