TypeScript-first agent coordination. Zero dependencies in core.
The open-source coordination library for autonomous agents. Every primitive you need to build production agents — typed, deterministic, and observable from the first run.
MIT licensed · TypeScript-first · zero core dependencies
Four invariants the whole SDK is built on.
Fail fast.
Errors are typed, attributed, and surfaced at the boundary where they occur. AgentResult<T> is a discriminated union — ok or err, never both, never neither. There is no swallowing an exception and retrying in secret. When something goes wrong, the SDK tells you which agent, which step, and why. Not after it cascades. Immediately.
Deterministic state transitions.
DeterministicStateRunner enforces a forward-only pipeline. Every state transition is logged before it executes. There is no rollback without an explicit decision. You know what step your pipeline is on at any moment because the SDK records every transition. Not eventually — at the point it happens.
Adapters all the way down.
LLMAdapter, VectorStoreAdapter, ToolAdapter — clean interfaces for every external dependency. Swap your LLM provider without touching agent code. Migrate your vector store without changing your retrieval logic. Vendor lock-in is a choice, not a default. The core has zero vendor dependencies.
Observability is the default.
StepAuditLog, PromptLayer, ObservabilityHooks, and ContextWindowManager ship with the SDK and activate on first run. You do not configure observability — you opt out of it. Every run produces a full audit trail. Every prompt is versioned. Every token cost is attributed to the step that incurred it.
Four layers, each building on the last.
AgentResult<T>, defineAgent, DeterministicStateRunner, StepAuditLog, PromptLayer, ContextWindowManager. The invariants the whole system is built on. Zero external dependencies.
LLMAdapter, VectorStoreAdapter, ToolAdapter interface definitions. No implementations — just the contracts. Implementations live in adapter packages or your own code.
Composable capability units that agents can declare and invoke. A skill wraps a tool or capability with a typed interface and policy-evaluable intent.
HitlAgent, SynthesisAgent, and the base agent primitives. Opinionated implementations built on core/ and adapters/. Use them directly or as reference implementations.
Two packages. One system.
The full coordination library. Core primitives, adapter interfaces, skills system, and first-party agents.
MCP bridge. Wrap any MCP server as an Agiorcx tool. Expose any Agiorcx agent as an MCP server.
Where we are. Where we are going.
- core/ primitives: AgentResult<T>, defineAgent, DeterministicStateRunner
- PromptLayer and StepAuditLog
- adapter interfaces: LLMAdapter, VectorStoreAdapter, ToolAdapter
- @agiorcx/mcp: MCPToolAdapter and AgentMCPServer
- First npm publish
- Python bindings
- SynthesisAgent — multi-source synthesis with attribution
- ContextWindowManager streaming strategy
- Orion reference implementation
- Stable API contract — no breaking changes without major version
- Full documentation site
- Platform console launch
- Cordax multi-agent coordination layer