Work
- 01
Self-hosted agent runtime
Building blocks for running a small team of agents on your own hardware. Each agent has a task inbox, a heartbeat and persistent memory. Code an agent writes runs in a sandbox with no network access and no secrets, outbound requests pass through a per-agent allowlist, messages between agents are signed, and new capabilities wait at an approval gate. Every action is recorded in an audit chain.
- 02
Kasumi
A privacy-first gateway between agents and model providers. It routes requests across providers, removes personal data before it leaves your machine, and keeps no prompts or responses by default.
- 03
Agent memory
Persistent memory for agents: facts, messages and logs stored in a searchable form, so an agent can recall earlier work across sessions. Each prompt gets a size-bounded, labeled block of the most relevant memory, not the whole history.
- 04
Cost-first routing
Kasumi's auto-router estimates how hard a request is and sends it to the cheapest model that can handle it, trying free tiers first and paid models last. Every request is counted per project (tokens, cost, model mix) without storing its content, and daily request caps keep a fleet of agents inside a set limit.
- 05
Long-run benchmarks
A harness that runs a fleet of autonomous agents in a persistent multiplayer game world, on an emulated server, for hours or days at a stretch. Agents set their own plans, gather, trade and train skills, and are scored on the game's leaderboards and on what they earn per unit of compute. Long runs show what short prompts hide: stalls, drift and runaway cost.