Status: 🚧 Active Research & Development (Prototype & Engineering Exploration)
Custos is an experimental, local-first development and execution environment engineered for autonomous agent workflows across software engineering, scientific inquiry, and automated research pipelines.
Rather than treating AI agents as opaque prompt loops or ephemeral chat completions, Custos introduces an evidence-bound, capability-gated architecture where human sovereignty is preserved and every external mutation is strictly auditable.
Core Architectural Pillars (In Development)
flowchart LR
Model[Untrusted AI Model] -->|Emits ActionIntent| Kernel[Custos Kernel]
Kernel -->|Signs Single-Use Permit| Gateway[Capability Gateway]
Gateway -->|Sandboxed Mutation| Outbox[Filesystem / Shell / Network]
Outbox -->|Generates Objective Diff / Trace| Evidence[Verifiable Artifact Store]
Task First-Class (Durable State Persistence):
- Work units live independently of transient client connections and chat sessions.
- Built on a local SQLite Write-Ahead Logging (WAL) engine ensuring crash-resilient command ledgers and idempotent workflow recovery.
Proposal $\neq$ Decision $\neq$ Effect (Strict Capability Gating):
- Untrusted model outputs emit declarative
ActionIntentspecifications rather than executing direct side-effects. - The security kernel validates intent against explicit policy gates before granting single-use execution permits.
- Untrusted model outputs emit declarative
Evidence-Backed Task Completion:
- Tasks cannot declare completion through conversational assertion alone.
- Acceptance criteria require verifiable empirical artifacts (test suite execution logs, deterministic AST diffs, citation audit trails).
Current Roadmap & Exploration
- Formalizing the sandboxed execution layer and sub-process jail on macOS / Linux.
- Benchmarking context compaction and multi-agent coordination protocols without token bloat.