COLOSSUS · ALPHA
Agent operations for mission environments¶
Colossus is an alpha-stage runtime for organizations that need AI agents to perform real work under explicit authority. Policy, approvals, sandbox boundaries, durable state, and audit evidence stay part of the operation whether the system is connected, local, or offline.
Alpha software means change is expected. Review upgrade and compatibility guidance before updating a deployment you depend on.
Operational assurance, not hidden autonomy¶
Colossus is being built for enterprise, public-sector, regulated, and disconnected environments where an agent action must be reviewable, bounded, and recoverable.
Account for every effect¶
Requested actions, policy decisions, approvals, execution, release, and uncertain outcomes produce durable evidence in a hash-chained journal.
Enforce the boundary¶
Strict schemas, one-use permits, sandbox profiles, resource ceilings, and quarantined output keep authority explicit from request through release.
Operate online or offline¶
Use hosted or local models, run a credential-free offline proof, and prepare controlled or air-gapped deployments without changing the core authorization path.
How Colossus governs work¶
flowchart LR
I["CLI · TUI · Desktop · SDK"] --> R["Durable runs<br/>agent · workflow · research"]
R --> C["Capability catalog<br/>strict tool schemas"]
C --> K["Safety Kernel<br/>policy · approval · one-use permit"]
K --> B["Execution boundary<br/>native · OCI · Windows · external"]
B --> A["Effect adapters<br/>models · files · processes · network · integrations"]
A --> Q["Quarantine and release<br/>bounded result"]
Q --> R
R --> J["Hash-chained journal<br/>audit · recovery · projections"]
K --> J
B --> J
Interfaces start durable work but do not own model, policy, tool, or state behavior. The runtime validates a requested capability, authorizes one exact effect, binds it to an execution boundary, and releases bounded output only after the required checks. Lifecycle evidence remains available for audit and recovery.
Choose your path¶
Start with the operational outcome you need.
Install it and prove the runtime offline before adding credentials.
Configure identity, access, isolation, storage, policy, and audit.
Add workflows, enterprise integrations, MCP servers, skills, and packs.
Use the application SDKs or work on the Rust runtime and Desktop.
Looking for exact commands, fields, schemas, formats, or limits? Open the Reference.