VSvarunsingla.com

← All entries

Day 46· · 3 min read

Multi-Agent Debugging in Production

Foundations & Protocols
Viral app of the day

Why it went viral: Google accidentally shipped COSMO to the Play Store on May 1 and pulled it within

hours -- but not before screenshots, APK extracts, and demos hit Reddit, X, and YouTube. By May 2 it was a top Google Trend. The leaked app reveals Google's answer to the OpenAI×Qualcomm×MediaTek What's notable architecturally: COSMO is built on Gemini Nano running on-device. It is proactive, not reactive -- it observes context and offers help without being asked. Six built-in agents (List Tracker, Document Writer, Calendar Event Suggester, Browser Agent, Deep Research, Quick Photo Lookup) plus on-device retrieval. This is the edge-inference stack from Day 40 hitting consumer scale through Android On-device models + a proactive agent loop + the OS as the surface = the next consumer AI form f

By the numbers
90%
of silent failures caught by 30-second verification
6
failure modes unique to agents
$0.14/M
DeepSeek V4 Flash -- new nano-tier

1. Why Agent Debugging Isn't APM

Traditional APM was built for deterministic services: a 500 means a 500. Agents fail in multi-turn, multi-tool sequences where the dashboard stays green and the user still gets a wrong answer. Six failure modes are unique to agents: tool misuse, context loss, goal drift, retry loops, cascading errors in multi-agent systems, and silent quality degradation. Of those, three are silent -- they require quality evaluation to

If your only signal is HTTP status codes, you are already blind to 60-70% of agent incidents. The dashboard must include trajectory scoring, output quality regressions, and per-step LLM-as-Judge verdicts.

BugWhat it looks like to the userOTEL signal that catches it
Silent failureAgent returns a confident but wrong answer. No exception, no 5xx. Sub-agent passed incomplete context to the next.Output-quality eval drops >5% on golden set. Trajectory score < 0.7. Trace pattern: tool span returns 200 but evaluator span scores low.
Context driftBy turn 4 the original constraint from turn 1 has faded. Phase transitions accumulate token noise.Cosine sim(goal, recent actions) < 0.7. Trace pattern: a span attribute agent.goal_alignment trending down across spans in the same trace.
Tool hallucinationAgent calls a tool that doesn't exist, or invents arguments. Error caught, then re-tried with a new fake.Tool error rate > 25% over 30s window OR 3 consecutive errors. Trace pattern: mcp.tool.error spans clustered + token usage spike 3×.

2. The 5 Hardest Bugs (and how they look in OTEL)

Silent failures, context drift, and quality degradation don't produce error codes. They produce plausible outputs. Latitude (Apr 2026): "30 seconds of verification catches 90% of silent failures" -- but only if you've wired evaluator spans, not just request spans, into your traces.

BugWhat it looks like to the userOTEL signal that catches it
Silent failureAgent returns a confident but wrong answer. No exception, no 5xx. Sub-agent passed incomplete context to the next.Output-quality eval drops >5% on golden set. Trajectory score < 0.7. Trace pattern: tool span returns 200 but evaluator span scores low.
Context driftBy turn 4 the original constraint from turn 1 has faded. Phase transitions accumulate token noise.Cosine sim(goal, recent actions) < 0.7. Trace pattern: a span attribute agent.goal_alignment trending down across spans in the same trace.
Tool hallucinationAgent calls a tool that doesn't exist, or invents arguments. Error caught, then re-tried with a new fake.Tool error rate > 25% over 30s window OR 3 consecutive errors. Trace pattern: mcp.tool.error spans clustered + token usage spike 3×.

3. The Hard Part: Correlating Traces Across Parallel Agents

When two sub-agents run concurrently and one poisons shared state, you need trace context propagated through every message boundary, not just function calls. OpenTelemetry's GenAI Semantic Conventions

Memory poisoningPersistent corruption: an unverified write to episodic memory infects future sessions. NIST CAISI Tier-1 risk.Vector-write spans without provenance attestation. Trace pattern: SVID mismatch on memory.write + retrieval score regression a day later.
Runaway cost loopLoop detection misses; agent calls the same tool with rolling-hash variants. $-burn alarm beats your latency alarm.Budget burn rate > 3× P95. Trace pattern: >3 same tool calls within rolling hash window OR token spend > 2× P95 in a 60s slice.

4. The 2026 Multi-Agent Debugging Stack

OpenTelemetry is now the lingua franca. Vendors fan out from a single OTEL pipeline. Pick by primary

PlatformWhen to reach for it
Arize PhoenixOSS, OTEL-native, OpenInference conventions, drift detection. Best free tier for teams already on OTEL.
LangfuseMIT-licensed, 21K+ stars, self-hostable. Strongest for EU AI Act data-residency. Now ships LLM-as-Judge evaluators in-product.
LangSmithLangGraph-native checkpoint rewind -- replay an agent from any point. Best if your stack is LangGraph 2.0.
BraintrustFull-lifecycle (eval + obs + regression). $80M raised. Best when prompts-as-code is mature.
New Relic SRE AgentAgentic root-cause analysis layered on top of traces. Auto-generates remediation hypotheses for the on-call.
Amazon OpenSearch unified obsMetrics + traces + agent debugging in one workspace. Native Prometheus + OTEL ingest. Strong for AWS-heavy estates.
Microsoft Agent 365 (May 1 GA)Not an APM, but the governance layer above it. Fleet registry, Entra-bound agent identity, real-time anomaly scoring, one-click kill switch. $15/user, or M365 E7 ($991).

5. The 3-Step Rollback Playbook

When a swarm starts misbehaving, the on-call has minutes -- not hours. The playbook every Agent SRE

PlatformWhen to reach for it
Arize PhoenixOSS, OTEL-native, OpenInference conventions, drift detection. Best free tier for teams already on OTEL.
LangfuseMIT-licensed, 21K+ stars, self-hostable. Strongest for EU AI Act data-residency. Now ships LLM-as-Judge evaluators in-product.
LangSmithLangGraph-native checkpoint rewind -- replay an agent from any point. Best if your stack is LangGraph 2.0.
BraintrustFull-lifecycle (eval + obs + regression). $80M raised. Best when prompts-as-code is mature.
New Relic SRE AgentAgentic root-cause analysis layered on top of traces. Auto-generates remediation hypotheses for the on-call.
Amazon OpenSearch unified obsMetrics + traces + agent debugging in one workspace. Native Prometheus + OTEL ingest. Strong for AWS-heavy estates.
Microsoft Agent 365 (May 1 GA)Not an APM, but the governance layer above it. Fleet registry, Entra-bound agent identity, real-time anomaly scoring, one-click kill switch. $15/user, or M365 E7 ($991).

6. Today's Breaking (May 1-4, 2026)

Step 2 -- Snapshot (T+30s to T+2m)Freeze working memory. Tag the current state as incident-yyyy-mm-dd-NN. Snapshot Redis L1, pgvector L2 episodic store, and Mem0 L3 unified API. Export the OTEL trace bundle for the last 15 minutes.Memory versioning + namespace isolation make this a single command (write-aside pattern, Day 23). The snapshot is what you replay against in dev to find the seed bug.
Step 3 -- Roll forward (T+2m to T+5m)Don't roll back to last week -- roll forward to a quarantined fork. Spin a canary off the snapshot, route 1% of traffic to the patched harness (e.g. tighter context-reset threshold, smaller MCP allowlist, narrower SVID scope). If the evaluator score recovers above the golden-set baseline, ramp 5% fi 25% fi 100% with auto-rollback wired to error rate +2% / cost +20% / latency +30%.This is the Day 23 5-stage CI/CD pipeline applied at incident speed. The trick: golden datasets and canary infra exist before the incident, not after.

7. Viral App of the Day -- Google COSMO

Why it went viral: Google accidentally shipped COSMO to the Play Store on May 1 and pulled it within hours -- but not before screenshots, APK extracts, and demos hit Reddit, X, and YouTube. By May 2 it was a top Google Trend. The leaked app reveals Google's answer to the OpenAI×Qualcomm×MediaTek

What's notable architecturally: COSMO is built on Gemini Nano running on-device. It is proactive, not reactive -- it observes context and offers help without being asked. Six built-in agents (List Tracker, Document Writer, Calendar Event Suggester, Browser Agent, Deep Research, Quick Photo Lookup) plus on-device retrieval. This is the edge-inference stack from Day 40 hitting consumer scale through Android

On-device models + a proactive agent loop + the OS as the surface = the next consumer AI form factor. Watch Google I/O on May 19. The agent-first phone is no longer a 2027 story; it's a Q3 2026 launch.

Practical takeaways
Tomorrow (Day 43): The Agent Eval Layer -- building golden d

Tomorrow (Day 43): The Agent Eval Layer -- building golden datasets that actually catch regressions, calibrating LLM-as-Judge against human labels, and the <2% regression CI gate that lets you ship safely on a Friday afternoon.

VS
Varun Singla
Singapore · About · Learning in public