What is Neusnap

Neusnap is a transaction layer for AI agents. It defines a strict execution boundary around multi-step agent workflows, ensuring that tool calls, model responses, and state writes either complete atomically or are fully rolled back.

What Neusnap is NOT

  • ×Not an agent framework
  • ×Not an LLM provider
  • ×Not a prompt orchestration tool
  • ×Not a UI or dashboard

Neusnap integrates with existing agent logic. It does not replace your agent framework.

Problem

Multi-step agent workflows fail mid-execution and leave corrupted state. A tool call executes and produces side effects. The LLM times out. Retry logic re-executes the tool call, duplicating the side effect. There is no atomic boundary around the workflow.

Neusnap Approach

Each agent run executes inside a transaction.

Tool calls and state writes are staged in a buffer.

LLM invocation occurs within the same boundary.

On success: all staged writes commit atomically.

On failure: all staged writes are discarded.

Retries start from a clean, deterministic state.

Architecture

1. Request
User or agent trigger initiates execution
2. Router
Policy evaluation selects model and fallback strategy
3. Model execution
OpenAI, Anthropic, or local model processes request
4. State buffer
Writes are staged, not yet visible externally
5. Commit or rollback
All writes finalize atomically or are fully discarded
6. Immutable output
Result is returned with guarantee applied

Where it sits in the stack

Your agent logic (LangChain, custom, etc.)
↓ Neusnap transaction layer
LLM APIs (OpenAI, Anthropic, local)
Tool execution (DBs, APIs, search)
State storage (Postgres, Redis, etc.)

Neusnap does not replace your agent framework. It enforces execution semantics beneath it.

Guarantees vs Non-Guarantees

Guaranteed
  • Atomic state writes
  • Deterministic retries
  • No duplicate tool execution on retry
  • Rollback on any failure inside transaction
Not Guaranteed
  • External API idempotency
  • LLM output determinism
  • Network latency
  • Zero failures

Neusnap guarantees execution semantics, not correctness of model output.