AI agents now write code, manage infrastructure and make decisions at scale. The bottleneck is no longer capability — it is accountability. AgentStateLabs builds the primitives that make agents governable: one content-addressed state graph, and six products on top of it.
Every one of them is a state problem. Today that state lives in chat logs, scratch files and someone's memory — none of which survive the session, and none of which you can audit.
What changed? A diff shows the edit, not the reasoning that produced it.
Why? The rationale was in a conversation nobody kept.
Who authorised it? Attribution stops at a service account.
Expected vs actual? Nothing recorded what the agent thought would happen.
Every product writes to the same content-addressed graph, so memory, code decisions, routing and test runs are branchable, blameable and auditable by default — not five tools you have to reconcile.
Content-addressed, branchable, blameable state. Every commit carries intent, reasoning, confidence and authority — so agent state reads like source history.
Rust core · 73 MCP tools · 8 bindings · 4 storage backends
Persistent, searchable, accountable memory. Write a fact once and every tool remembers it — with plans, branches, taint and provenance on top.
90+ MCP tools · REST API · Lens UI · SQLite or Postgres
A decision ledger, effect declarations and a call graph for every function — queryable by the agents writing the code, and checked into git beside it.
Rust core · 64 MCP tools · 9 languages · git-native sidecar
Threaded agent conversation and provenance built directly on the AgentStateGraph substrate, so a chat history is an auditable graph rather than a transcript.
Currently v0.3.5-beta.2
Run one scenario against many agents, capture every decision as an auditable graph, and score the runs side by side with a judge agent.
Scenario runner · judge scoring · run diffing
The Agent Execution Optimization Engine. Routes each task by cost-per-unit-of-work and your own agent history — not by token price alone.
Currently v0.1.0-alpha · suggester · proxy · embedded
CTXone holds session memory and plans. AgentStateDeveloper holds codebase context. AgentStateRouter decides which model runs the work. AgentStateCrucible proves the agent behaved. ThreadWeaver keeps the conversation. All of it stored in AgentStateGraph — all of it auditable.
AgentStateGraph ← the substrate everything else runs on
├── CTXone ← memory, plans, branches, provenance
├── AgentStateDeveloper ← symbols, effects, decision ledger
├── ThreadWeaver ← threaded conversation state
├── AgentStateRouter ← which agent, at what cost
└── AgentStateCrucible ← test runs, judge scoring, auditEvery product is self-hosted and MCP-native, so any MCP-compatible agent connects immediately. Adopt one, or adopt the graph underneath them all.