Rifft
Summary
A multi-agent pipeline fails at 2am and the log gives you a KeyError at the last step — but the actual break happened three agents back, in a schema field nobody validated. Rifft exists to close that gap between where the error surfaces and where the state actually went wrong.
Rifft is a passive debugging layer for production agent pipelines built on CrewAI, AutoGen, LangGraph, and similar frameworks. Drop in one import, wrap your entry point, and Rifft automatically captures handoffs, tool calls, and state mutations across every span. When a run fails, it walks the trace backwards to the first bad state — classifying the failure against the MAST taxonomy — and lets you replay from that exact handoff with patched inputs, without restarting the full crew. The side-by-side diff between the broken run and the fixed replay is where debugging time actually disappears. The ceiling arrives when your pipeline runs outside the supported frameworks or when you need on-premise trace storage.
Bottom line: Reach for Rifft when your CrewAI or LangGraph pipeline is failing in production and you need to find the bad handoff in minutes rather than hours — but if your compliance posture requires traces to never leave your infrastructure, there is no self-hosted option to fall back on.
Pricing Plans
SubscriptionLast verified 1 week ago- Price
- $49/month
- Free Tier
- 50K spans/month, 14-day retention, Causal graph & MAST classification, 1 workspace
FREE
For individual developers exploring Rifft.
- 50K spans/month
- 14-day retention
- Causal graph & MAST classification
- 1 workspace
PRO
For teams running agents in production.
- 500K spans/month
- 90-day retention
- Fork mode & replay
- Unlimited team members
- Slack & email alerts
- NL failure explanations
SCALE
For high-volume pipelines and larger orgs.
- 2M spans/month
- 1-year retention
- Everything in Pro
- Automatic regression detection
- Priority support
- $5 per 100K spans above 2M
View full pricing on rifft.dev →
Pricing may have changed since last verified. Check the official site for current plans.
Community Performance Report Card
No community ratings yet. Be the first to rate this tool!
Community Benchmarks Community
Sign in to submit a benchmarkNo community benchmarks yet. Be the first to share a real-world data point.
Pros
Sign in to edit- Backwards trace walking from error to root cause, so you identify the span that produced bad state instead of reading 12,000 tokens of log output in sequence.
- MAST failure classification across four categories, which means a handoff schema mismatch or a tool loop is recognized and labeled on first sight rather than diagnosed from scratch each time.
- Replay from any span with patched inputs — without rerunning the full crew from the start — so a fix hypothesis is confirmed in seconds rather than minutes of re-execution.
- Provider-agnostic instrumentation across seven-plus frameworks via a single import and decorator, so teams do not rewrite observability code when switching between CrewAI and LangGraph.
- Similar-run surfacing groups failures with prior runs that matched the same MAST class, so recurring agent bugs are visible as patterns before they accumulate into an incident.
Cons
Sign in to edit- Traces are stored on Rifft's cloud infrastructure — there is no self-hosted deployment path, no installable container, and no on-premise option listed. Teams whose contracts prohibit external telemetry data or whose compliance frameworks require data residency switch to OpenTelemetry-compatible self-hosted stacks (Jaeger, Langfuse self-hosted) even when the debugging experience is materially worse.
- The MAST taxonomy covers four classified failure classes. Agent failure modes outside those classes — for example, semantic drift in long-running conversations, reward hacking in tool selection, or cross-session memory corruption — are captured as raw spans but receive no classification or pattern-matching, leaving the debugging experience identical to unstructured log review.
- Rifft is a passive observer: it captures what happens but does not add validation gates, schema enforcement, or retry logic to the pipeline. Teams that want to prevent the failure class — not just diagnose it after the fact — build guardrails separately, maintaining a second layer of tooling alongside Rifft.
Community Reviews
Sign in to write a reviewNo reviews yet. Be the first to share your experience.
About
- API Available
- Yes
- Self-Hosted
- No
- Last Updated
- 2026-06-26T02:23:52.389Z
Best For
Who it's for
- Teams running production agent pipelines
- Developers using CrewAI, AutoGen or LangGraph
- Debugging complex multi-step agent interactions
What it does well
- Debugging failed multi-agent workflows
- Classifying failure modes with MAST taxonomy
- Replaying from specific handoffs without full restarts
- Comparing traces side-by-side after fixes
Integrations
Discussion Community
Sign in to commentNo discussion yet. Sign in to start the conversation.
Compare Rifft
Spotted incorrect or missing data? Join our community of contributors.
Sign Up to ContributeCommunity Notes & Tips Community
Sign in to contributeBe the first to contribute. General notes, observations, gotchas, and tips from people who use this tool day-to-day.
Frequently Asked Questions
- Is Rifft free?
- Rifft has a permanent free tier alongside paid upgrades (paid plans from $49/month). You can keep using a baseline version indefinitely without paying.
- Is Rifft open source?
- No — Rifft is a closed-source tool. Source code is not publicly available.
- Does Rifft have an API?
- Yes. Rifft exposes a developer API. See the official documentation at https://rifft.dev for details.
Hours Saved & ROI Stories Community
Sign in to contributeBe the first to contribute. Concrete time/cost savings, with context. e.g. "Cut my code review backlog from 4h to 45m per week."
Curated lists that include this category
Agent pipelines fail silently. An agent returns the wrong schema, the next agent parses what it can, and three steps later you get a KeyError that looks like a publisher bug — not the researcher’s unvalidated citation from four minutes earlier. Rifft instruments that chain with a single import and decorator, capturing every handoff, tool call, and state change as a span. When the run fails, Rifft walks backwards through the spans to surface the root cause — not the last error thrown, but the first place the state diverged from what downstream agents expected.
The differentiating mechanism is the MAST taxonomy: four classified failure modes (unverified output, tool loops, handoff schema mismatches, and context overflow) that Rifft maps each broken run to. This classification matters in practice because it converts a unique-looking stack trace into a category you have seen before, which also lets Rifft surface similar prior runs. When the same researcher agent passed a bad citation three times last week, you see that — before you spend an hour treating Tuesday’s failure as novel.
Rifft fits teams running agent pipelines that are already in production and failing in ways that raw logs cannot diagnose in reasonable time. The vendor states support for CrewAI, AutoGen, LangGraph, LlamaIndex, OpenAI Swarm, and custom stacks via OpenTelemetry. The integration path is five lines of code — one import, one decorator on the entry point — and the vendor states no exporters, collectors, or YAML configuration are required. Where Rifft breaks down: the tool is a cloud-hosted service with no self-hosted deployment path, so traces leave your infrastructure. Teams with data residency requirements or contracts that prohibit external telemetry have no workaround within the product.
The replay mechanic — forking a run from any span with patched inputs or a swapped agent — is the sharpest production-facing feature. The vendor reports 2.8 seconds end-to-end replay time, which means iterating on a fix and confirming it resolves the broken handoff takes less time than re-reading the original trace. Side-by-side run diffing is available to compare the failed run against the passing replay before promoting the fix.