Agent SDK
A deterministic TypeScript agent runtime for planning, reasoning, memory, and multi-agent orchestration. Native providers for OpenAI, Anthropic, Azure, Bedrock, and Vertex — no LangChain required — with a zero-graph, message-first loop plus built-in summarisation, guardrails, structured output, and tracing.
What's in the box
Teams building custom agent logic in TypeScript and needing deterministic execution, planning, reasoning, and human-in-the-loop control. Each capability is opt-in — use the parts that fit, leave the rest.
Native LLM Providers
Direct bindings for OpenAI, Anthropic, Azure, Bedrock, Vertex, and any OpenAI-compatible endpoint — with normalised token accounting and no LangChain dependency.
Message-First Runtime
Executes agent reasoning through a transparent, message-driven loop with full conversational traceability.
Reasoning & Reflection
Provider-native reasoning (OpenAI Responses, Anthropic thinking, Vertex thoughts) paired with post-tool reflection you can route to memory or the plan.
Planning (TODO)
Structures agent workflows through a managed TODO plan with configurable replanning on failure.
Pause / Resume & HITL
Snapshot and restore agent state, gate sensitive tools behind approval, and ask the user structured questions mid-run.
Summarisation
Compresses long context with token-aware rewriting while keeping full tool payloads retrievable on demand.
Guardrails
Applies runtime safety and policy controls to govern agent behaviour and outputs.
Structured Output
Supply a Zod schema and the runtime returns validated, fully typed JSON via a hidden finalize tool.
Multi-Agent & Tracing
Compose agents as tools or explicit handoffs, with structured traces streamed to file, HTTP, OTLP, or Cognipeer sinks.
How it runs
A single deterministic loop: plan → call tools → reflect → checkpoint. No graph DSL to learn, no implicit globals. Everything you would inspect in a debugger is on the result.
Inputs
- User messages
- Tool registry
- Reasoning profile
- Approval policies
State
state.planstate.reflectionsstate.historystate.snapshot
Side effects
- Tool calls
- Tracing events
- Approval requests
- Token accounting
Quickstart
Install, configure, run. The example below is the smallest piece of code that does something useful in production.
How it compares
Against the flagship · runtime options teams most often weigh — focused on operational concerns, not feature inventories.
| Capability | Agent SDK | LangChain (JS) | LangGraph | OpenAI Assistants API |
|---|---|---|---|---|
| Deterministic Runtime | ● native | ○ missing | ○ missing | ○ missing |
| Planning | ● native | ◐ partial | ◐ partial | ○ missing |
| Summarisation | ● native | ○ missing | ○ missing | ○ missing |
| Guardrails | ● native | ○ missing | ○ missing | ◐ partial |
| Interrupt | ● native | ○ missing | ○ missing | ○ missing |
| Pause / Continue | ● native | ○ missing | ◐ partial | ○ missing |
Patterns that use Agent SDK
Production-tested recipes powered by this package.
SRE Incident Response Agent
Build an SRE agent that reads Confluence runbooks, ingests them into Console RAG, searches vector knowledge during incidents, and posts Jira updates through tools.
Agent Development With Tracing
Develop custom agents with Agent SDK and feed execution traces into Console so teams can compare runs, inspect tool behavior, and debug workflows faster.
Autonomous AI Agents
Build agents that plan, execute tools, and deliver structured outputs with full deterministic control and observability.