Agent reliability · Open source · 2026
EvidenceForge
An open-source framework for building auditable, evidence-citing AI agents.
EvidenceForge starts from a simple rule: a claim is not a result. An agent run is only finished when each material claim cites the record, document, or tool output it came from.
The framework keeps a free core and a clear path to commercial compliance use, so a team can adopt the discipline first and pay only when it needs the governed deployment around it.
In practice that means fewer arguments about what an agent did, and a run a second person can audit without re-running the whole system on trust.
More work
API tooling · Developer product
mimicplus
Record a HAR, get a versioned spec, a typed Python client, OpenAPI, and a breaking-change check.
Multi-agent systems
Governed Agent Engine
A governed multi-agent engine: Termux-installable, stdlib plus PyYAML, with spec binding and audit.
Audit infrastructure
Verity
An append-only evidence ledger for agent runs. Claims must cite evidence.