If you run more than one AI agent in production, you've already hit the observability gap.
Single-agent tracing is a solved problem. Langfuse, LangSmith, Arize Phoenix, and a dozen others will show you exactly what happened inside one agent's LLM calls. That's table stakes in 2026.
But when Agent C errors because Agent B sent it bad data, which itself came from Agent A's failed tool call, most of these tools give you three separate error logs with no connection between them.
Here's where the landscape actually stands.
The Per-Agent Tracers
Langfuse (22.9k GitHub stars, acquired by ClickHouse in January 2026) is the open-source default. It traces LLM calls, manages prompts, runs evaluations. 2,000+ paying customers. 19 of the Fortune 50 use it. The ClickHouse acquisition gives it serious infrastructure backing. What it doesn't do: correlate errors across agents that share no direct function call.
LangSmith is LangChain's proprietary tracing platform. Deep integration with the LangChain ecosystem. Automatic pattern clustering, failure mode detection, online evals. Pricing starts free (5k traces/month), then $39/seat + $2.50 per 1k traces. Strong if you're already in the LangChain world. Same limitation: per-agent scope, no fleet-wide cascade view.
AgentOps (3.1k stars) has the lowest performance overhead in 2026 benchmarks at 12%. Supports CrewAI, Agno, OpenAI Agents SDK, LangGraph, AutoGen. Good multi-framework story. Session replays are useful for debugging single agent runs.
Arize Phoenix (open source + enterprise) comes from ML observability. Drift detection, performance degradation monitoring, agent graph visualization. Recently added Amazon Bedrock Agents support. Their trajectory mapping can detect recursive patterns within one agent's execution.
RagaAI Catalyst is Python-only, self-hosted, with 50+ custom metrics and execution graph visualization. Designed for multi-agent systems but focused on tracing rather than cross-agent failure attribution.
The Enterprise Entrants
Braintrust raised $80M Series B in February 2026 at an $800M valuation. Notion, Replit, Cloudflare, Ramp, Dropbox are customers. They cluster production traces into "Topics" and use AI to suggest improvements. The clustering approach groups similar failures but doesn't show the causal chain between agents.
Datadog added agent monitoring to its LLM Observability suite. Decision-path graphs, tool invocation tracking, infinite loop detection, cost/token analysis. Integrated with Google's Agent Development Kit. Datadog's strength is correlation with existing APM data. If your agents run on infrastructure you already monitor with Datadog, you get agent + infra correlation for free.
Splunk went GA in February 2026 with agent monitoring. Traces and maps agent workflows with dependency analysis. Integrated with Cisco AI Defense for PII detection and prompt injection. Security-focused, SOC-ready. Enterprise pricing.
New Relic launched its agent platform in February 2026 with OpenTelemetry integration and a no-code agentic builder. Targets teams already using New Relic for APM.
The Cascade Detection Gap
Here's what's interesting. Everyone I listed above traces what happened inside individual agents. Some can even show you a graph of how agents communicated. But when a failure propagates across autonomous agents, amplifies through feedback loops, and compounds into a system-wide problem before a human can react, most of these tools give you symptoms, not root causes.
OWASP formalized this in ASI08 (December 2025): cascading failures in agentic AI. Their definition: "A cascading failure occurs when a single fault propagates across autonomous agents and compounds into system-wide harm." Their recommended mitigation includes "comprehensive observability with automated cascade pattern detection."
Galileo is the closest to addressing this. Their Signals tool uses ML-based detection (their Luna-2 evaluation models) to find when malformed tool output corrupts working memory and propagates through decision chains. It's semantic failure detection. The limitation: ML detection tells you THAT something happened. It doesn't show you the evidence at each hop.
Maxim AI ($3M seed, backed by Postman and Chargebee founders) takes an end-to-end approach with agent simulation, evaluation, and production observability. One of five platforms identified as leading in 2026, but cascade-specific features are still emerging.
Where It's Headed
The pattern is clear: single-agent tracing is commoditized. The next battleground is multi-agent correlation, specifically:
Cascade replay with evidence: Not just "Agent C failed" but "here's the payload Agent A returned, here's how Agent B transformed it, here's why Agent C broke." Deterministic replay with stored payloads at each hop.
Fleet-wide health: A single dashboard showing all your agents, their last heartbeat, uptime percentage, active warnings, and which agent is currently degrading the fleet.
Alert deduplication: The same alert from the same agent 19 times in an hour is noise. Smart escalation (info -> warning -> critical based on recurrence) is needed.
This is exactly what we're building with AgentWatch. It's a TypeScript library (npm install @nicofains1/agentwatch) that stores evidence at each hop and walks the trace DAG backward when something breaks. Deterministic replay, not ML pattern detection. Self-hosted with SQLite, zero external dependencies.
We built it because we run 7 AI agents on cron schedules, and we kept hitting the same problem: one agent fails, three others break, and we'd spend hours reading separate logs trying to figure out the chain.
The 0.95^10 problem is real. If each of your 10 agents is 95% reliable individually, your system is only 60% reliable. The 40% that goes wrong needs better tooling than per-agent traces.
Quick Reference
| Tool | Type | Cascade Detection | Pricing |
|---|---|---|---|
| Langfuse | OSS | No (per-agent) | Free + Cloud |
| LangSmith | Proprietary | No (per-agent) | Free tier + $39/seat |
| AgentOps | Commercial | Limited | Undisclosed |
| Arize Phoenix | OSS + Enterprise | Recursive pattern (single agent) | Free + Enterprise |
| RagaAI Catalyst | OSS | No (Python-only) | Free |
| Braintrust | Proprietary | Topic clustering | Enterprise |
| Galileo | Proprietary | ML-based detection | Undisclosed |
| Datadog | Enterprise | Dependency correlation | $0.40-0.60/GB |
| Splunk | Enterprise | Workflow mapping | Enterprise |
| AgentWatch | OSS | Deterministic cascade replay | Free (self-hosted) |
The tools that exist are good at what they do. They just weren't built for the problem that matters most when you scale past a single agent.
Top comments (0)