Anchor — the Durable Execution Engine for Mission-Critical AI Agents

Description and use

Anchor delivers crash‑proof execution for multi‑step AI agents by atomically journaling side‑effects in PostgreSQL, fencing epochs, and providing a human‑review queue for unsafe tools. It eliminates duplicate token costs, double billing, and high infrastructure taxes.

Problem

AI agents executing multi‑step workflows crash due to OOM, redeploys, spot preemption, or network drops, causing duplicate side‑effects, lost state, and wasted token costs.

Outcome

Anchor reduces duplicate token costs by >$12k/month, eliminates double‑billing and spam, and removes the $5k+ infrastructure tax of external orchestrators. It achieves sub‑second recovery and guarantees idempotent side‑effects.

Workflow

• 1. Install Anchor: pip install anchor-runtime anchor dev
• 2. Define tools with @anchor.tool (specify safety and idempotency)
• 3. Define an agent with @anchor.agent and yield anchor.ToolCall for each step
• 4. Run the agent; Anchor writes TOOL_INTENT and TOOL_RESULT entries to PostgreSQL
• 5. On crash, a secondary worker acquires the lease, checks the journal, and replays completed steps
• 6. If a tool was unsafe and crashed mid‑execution, Anchor sets run to needs_review; operators resolve via mark_executed or mark_not_executed

Evidence

Unmanaged retries waste over $12,400/month in duplicate prompt tokens

Source

Anchor atomically writes a TOOL_INTENT journal entry to PostgreSQL

Source

If a worker process dies mid‑execution, the secondary worker checks the journal on recovery—replaying completed steps in <5ms

Source