A calendar API built for AI agents.

Commit a booking. The endpoint accepts a slot and an agent key, runs the conflict engine, and either writes the booking or tells the agent who outranked it.

# pip install agentdraft
from agentdraft import Client, Conflict

agent = Client(api_key="avs_live_...")

try:
    booking = agent.bookings.create(
        user_id="u_abc",
        start="2026-06-01T14:00:00Z",
        end="2026-06-01T14:30:00Z",
        title="Intro call",
    )
    # 201 COMMITTED — exactly one agent reaches here per slot
except Conflict as c:
    # 409 — c.winning_agent / c.winning_priority / c.audit_id
    # fall back: pick another slot, ask the user, or wait.
    ...

TypeScript, REST, LangChain, and MCP examples are in the docs.

Calendar APIs were designed for one writer at a time. Google Calendar, Microsoft Graph, CalDAV — none of them expose a slot-level compare-and-swap. An agent's "check availability" and an agent's "commit booking" are two separate network calls, and other agents can complete their own check-and-write between them. Two agents both pass the availability check; both succeed in writing; the human ends up with overlapping events. We call that a multi-agent calendar collision and it gets more common every month.

AgentDraft is a thin coordination layer in front of whichever calendar your agents already write to. The atomic guarantee lives at the storage layer (DynamoDB conditional writes), not in application code, so the check is the write. Every booking is a single TransactWriteItems across one row per time bucket — exactly one writer wins per slot, every loser gets a typed 409 carrying the winner's identity, priority, and audit reference.

Full reference in the docs; the formal contract is on the protocol spec page.

agent-1 ────┐
agent-2 ────┼──▶  AgentDraft  ──▶  Google / M365 / CalDAV
agent-3 ────┘    (coordinate)     (store)

Every agent that needs to write to a user's calendar calls AgentDraft first. The conflict engine resolves the race; exactly one agent gets 201 COMMITTED back. That agent's write propagates to the underlying calendar through whichever connector the user is on. The losing agents never reach the calendar at all, so the availability surfaces stay coherent.

• How a deterministic conflict engine resolves 8,217 collisions — the math behind the engine, with p99 latency.
• Why AI scheduling agents collide — and what a coordination layer looks like — the protocol-level case for AgentDraft.
• Multi-agent calendar collision — definition — the failure mode, formally.