Case studies
Proof in production.
Real systems we design, run, or open-source. Enough detail to be useful, without naming client hosts unless they approve a public write-up.
Each study maps back to the consulting stack: operable agents, durable context, and intentional tokens. New to the vocabulary? Start with How Agents Work.
Gridcraft: RTS command for hypervisors and agents
StarCraft-style map of Proxmox bases, VMs, and apps — frontier models as SCVs that inspect context and dispatch work. Live demo included.
- Harness
- Ops
- Playable
The management layer
When tight integration stops scaling: self-supervision, lock-in, mixed fleets, and governance — and why we split execution from management in JARVIS v4.
- Agents
- Context
- Governance
One context window per server
The fleet's context is 4.3× a single 200K window. Binding each context to one SSH host. Hibernation as paging is what let it reach 164 agents on one laptop.
- Context
- Tokens
- Scale
Agent OS: the directory is the program
JARVIS v4 as a Unix-like agent OS: tmux processes, static-linked apps, dual sandbox files, and why ls is still the fleet audit tool.
- Agents
- Context
- Ops
JARVIS v4: the harness we run on ourselves
Domain-scoped agents, Rocket.Chat dispatch, tmux sessions, SSH sandboxes, and a 1:1:1:1 naming scheme, MIT-licensed and in daily use.
- Harness
- Agents
- Open source
Want the concepts without the architecture deep-dive?
More studies land here as we publish them. Client work stays under NDA until they sign off on a public version.