Open source · MIT licence

Debug multi-agent AI
where it actually breaks

Your multi-agent pipeline failed. You don't know which agent caused it, why it made that decision, or how the failure spread. Rifft shows you.


The problem

Multi-agent failures are hard to debug

Agent A passes bad data to Agent B. Agent B treats it as fact. By the time Agent C fails, you're three levels deep with no idea where it started. Standard logs tell you what happened. Rifft tells you why.

orchestratorstarted
researcherexecuting task...
researcherpassed output to writer
writerfailed — but why?
rifft detected root cause:
researcherunverified output propagation

Three things Rifft gives you that nothing else does

01

See exactly which agent caused the cascade

Rifft traces the failure back to its origin — not just where the pipeline crashed, but which agent's decision started the chain reaction.

02

Understand why the agent made that decision

Every agent decision is captured with its full context — the prompt state, conversation history, and tools available at the moment it chose.

03

Replay and fix it without re-running everything

Step through any trace, edit the agent's state at any point, and fork to test whether your fix would have worked — without paying for another full run.


Running in 5 minutes

crewai_example.py
1import rifft
2import rifft.adapters.crewai # one line
3
4rifft.init(
5 project_id="my-project",
6 endpoint="http://localhost:4318"
7)
8
9# your existing crew, unchanged
10crew = Crew(agents=[...], tasks=[...])
11result = crew.kickoff()
12
13# open localhost:3000 to see the trace
01

Start the stack

git clone https://github.com/rifft-dev/rifft then docker compose up. Everything starts together.

02

Install the SDK

pip install rifft rifft-crewai. Also works with AutoGen: rifft-autogen.

03

Add one import

No decorators, no code changes to your agents. The adapter instruments everything on import.

04

Open the debugger

localhost:3000 shows the communication graph, failure classification, and replay controls.


Stop guessing why your agents fail

Open source, self-hostable, MIT licence. Running in 5 minutes.