Home
Three guarantees¶
-
No hidden prompts
Every model call consumes inspectable, typed
Messageobjects. No prompt is ever constructed outside your view. -
No runaway loops
Guards are enforced inside the kernel, every iteration — not bolted on. Step, cost, token, wall-clock, and delegation ceilings are structural.
-
Deterministic replay
Every step emits a trace event; any run can be recorded and replayed step-for-step without calling the provider or any tool.
What's in the box¶
| Plane | Ships with |
|---|---|
| Kernel | step loop, explicit serializable state, guards, tracer, HITL interrupt/resume, parallel tools, cancellation, sub-agents |
| Middleware | retry, fallback, loop guard, structured output, compaction, cost tracking, caching, guardrails (PII/injection/policy), circuit breaker, idempotency, rate limit, deterministic replay, memory recall, tenant budgets, tool timeout/truncation/sandbox |
| Backends | checkpoints (in-memory, SQLite, Redis, Postgres); memory (vector, buffer, pgvector) with pluggable embedders |
| Providers | OpenAI, Anthropic — and any OpenAI-compatible endpoint (Ollama, vLLM, Groq, …). See Models |
| Adapters | MCP (tools), A2A (remote agents), OpenTelemetry (spans) |
| Tooling | CLI (init/run/chat/dev/trace/eval/doctor/plugin), eval harness, orchestration patterns |
Why Spine¶
The integration layer that made monolithic frameworks valuable has been commoditized by open standards (MCP for tools, A2A for agents). The durable value sits in the reliability runtime — the execution loop, guards, durable state, and observability that decide whether an agent survives production. Spine owns that core and lets the standards do the integration.