waypoint-sdk 0.1.0

A lightweight Python SDK for building fault-tolerant LLM agent workflows. It enables agent systems to recover from crashes by replaying execution from checkpoints, without re-invoking deterministic operations like LLM calls or completed tool invocations.

Waypoint

A Python SDK for making LLM agent workflows fault-tolerant via event sourcing.

When an agent crashes mid-execution, Waypoint lets you resume from the last successful step—without re-running LLM calls or tool invocations that already completed. It does this by logging every step's input/output to an append-only PostgreSQL journal, then replaying from checkpoints on recovery.

Getting Started

Prerequisites

• Docker + Docker Compose
• uv (Python 3.13+)

Clone & Start

git clone [e-mail]:aybruhm/waypoint.git
cd waypoint
make up

This starts the API gateway on http://localhost:9654 and PostgreSQL. The gateway auto-reloads on code changes.

Run Migrations

make run_migrations

Run Examples

# 3-step agent, no LLM
uv run python -m sdk.examples.simple_agent

# Mocked LLM + crash recovery demo
uv run python -m sdk.examples.agent_with_llm_mock

Stop

make down

Makefile Reference

Command	Description
make up / make start	Build & start containers (detached)
make down / make stop	Stop & remove containers
make run_migrations	Apply pending Alembic migrations
make revert_migrations	Roll back last migration
make add_migration MSG="msg"	Auto-generate new migration
make show_current_db_head	Show current migration version
make show_db_heads	List all migration heads

---

What It Solves

LLM agent crashes create three problems:

1. Wasted spend: LLM calls that succeeded before the crash get re-invoked on retry.
2. Lost context: No record of what happened, what state the agent was in, or which step failed.
3. Duplicate effects: Retrying a tool call (e.g., an API write) can create duplicates or break idempotency.

Waypoint avoids all three by persisting every step's result. On crash, you resume from the checkpoint—cached LLM responses return instantly, tool outputs are reused, and execution continues from the next step.

---

Architecture

Agent Code
    ↓
@checkpoint decorators (Waypoint SDK)
    ↓
┌────────────────┬─────────────────┬──────────────────┐
│ Event Journal  │ Checkpoint Mgr  │ Replay Engine    │
│ (append-only)  │ (progress)      │ (deterministic)  │
└────────────────┴─────────────────┴──────────────────┘
    ↓
PostgreSQL

---

Core Concepts

Concept	Description
Execution	A single run of an agent workflow, identified by a UUID.
Step	A decorated async function (@checkpoint("name")). Each step runs once per execution.
Checkpoint	A persisted record of a step's input/output + execution position.
Event Journal	Append-only log of all steps across all executions (PostgreSQL).
Replay	Reconstructing state by reading checkpoints in order, skipping re-execution.

---

How It Works

@checkpoint("step_name")
async def my_step(input):
    return output

The decorator:

1. Checks if a checkpoint exists for this step in the current execution.
2. If yes: returns cached output immediately (no function execution).
3. If no: runs the function, persists input/output as a checkpoint, returns output.

On crash, create a new Waypoint instance and call resume(execution_id). The SDK rebuilds state from the journal and continues from the next uncompleted step.

---

Key Properties

• Deterministic replay — Same inputs always produce same outputs; no re-execution.
• LLM call caching — Cached responses are returned on replay (zero token cost).
• Framework-agnostic — Works with LangChain, CrewAI, custom async agents, FastAPI, etc.
• Minimal integration — Add @checkpoint decorators (one per step). ~3 lines of change per step.
• Full history — Query every step, error, and state transition by execution ID.

---

When to Use

• Long-running agent workflows (minutes to hours) where crashes are expensive.
• Cost-sensitive apps where re-calling LLMs on retry is unacceptable.
• Teams needing audit trails for agent behavior and debugging.
• Agent-as-a-service platforms running untrusted/user-submitted agents.

---

When Not to Use (Next Steps)

• Distributed/multi-machine workflows (Waypoint is single-process).
• High-throughput task queues (use Celery, Temporal, etc.).
• Simple chatbots with no multi-step orchestration.

---

Stack

• Python 3.13+
• asyncio
• FastAPI (gateway demo only; SDK is framework-agnostic)
• PostgreSQL (events + checkpoints)
• Pydantic + JSON serialization