Plans
Goal decomposition. Up to three reflection cycles before a plan ships. Context carries forward across runs.
Six primitives. The same vocabulary infrastructure software has used for forty years — applied to AI agents that finish work.
Goal decomposition. Up to three reflection cycles before a plan ships. Context carries forward across runs.
Six step types: llm_call, tool_call, sub_agent, human_approval, wait, webhook. Dependencies resolved automatically.
Durable async queue. Survives serverless cold starts. Auto-retry up to three times. Claimed-by tracking.
Append-only log. Tokens, duration, output — every action recorded. Feeds the realtime constellation telemetry.
Persistent KV per (org, agent). Optional TTL. Vector search over knowledge documents via pgvector + HNSW.
Goal-achievement evaluation between runs. Auto plan revision when reflection finds a gap. Human gate when needed.
Every plan, every task, every memory — alive in one canvas.
Goal decomposition. Up to 3 reflection cycles before ship.
{ "goal": "...", "reflection_cycles": 3 }Ordered DAG. Six step types. Dependencies resolved.
{ "type": "llm_call", "depends_on": [1, 3] }Durable async queue. Survives cold starts. Auto-retry × 3.
{ "status": "queued", "retries": 0 }Append-only log. Tokens, duration, output — every action.
{ "tokens": 412, "duration_ms": 1840 }Persistent KV per (org, agent). Optional TTL. Vector search.
{ "key": "client.prefs", "ttl_seconds": 2592000 }Gmail, Drive, Calendar, Jobber, Stripe, OpenAI, Anthropic, xAI, and arbitrary webhooks — all live as first-class context inside every agent.
Stand up your first plan in under three minutes. No credit card. 500 free credits. Six agents. One constellation.
Arcasya / V2.1 / Operator-grade orchestration