NewAgent Observability — find failures before customers do.

Observability built for agents

AI can't debug what it can't see. Give agents full production context to debug applications, infrastructure, and other agents.Built on open standards. Object storage for telemetry. Serverless compute for queries. Sub‑second agentic queries at 80% lower cost.
Go Live in 15 minutes

900M+agent spans/month with zero sampling3PB+production telemetry ingested/month

    • Fello
    • WisdomAI

Largest customer deployment

0TB+logs/day
0M+time series/hour
0Magent spans/day
<0sP99 query latency

built by engineers behind Rubrik · Amazon S3 · DynamoDB · Snowflake

Customer stories

Ashish Gupta

Agent observability

3M+ agent traces/day

“Having both agent observability and full-stack observability in one place makes debugging faster and easier. We're catching issues before our customers do, with full visibility and no sampling across millions of agent traces.”

Fello

Ashish Gupta

Principal Engineer

Read case study

The real problem

AI changed debugging.
Observability architecture didn't.

Disk-backed, fixed-compute stacks were built for dashboards, not agents at scale.

So you sample, cut retention, or overprovision capacity.

Sampling → blind spots

Agents are non-deterministic, so you can't predict which traces matter. Sample them, and rare failures disappear before you know to look.

Expensive at full fidelity

1,000× query overload

Fixed clusters were sized for a few humans, not thousands of concurrent agent questions.

Unreliable · timeouts · capacity spikes

>1s stalls the agent

Agents generate 100–1,000× more telemetry. Human-scale query latency is too slow for automated debugging loops.

Observability becomes the bottleneck

Software + agents

Observability is becoming headless

1 · Your software

Agents debug production software.

The same telemetry your engineers already rely on, now queryable by AI.

Logs·Metrics·Traces·APM·RUM·Synthetics

Explore full-stack observability

2 · Your agents

Agents debug other agents.

Find what failed, prove the fix, and keep improving the agent in production.

Traces·Evals·Prompts·Insights

Explore agent observability

3 · Full context

To really debug the agent, you need both.

Infrastructure, applications, and the agent itself, so AI never debugs half-blind.

trace · api-gateway post /api/v1/chat

Duration

29.78s

Spans

151

Services

4

Errors

9

Est. cost

$0.0005

Tokens

4,224

InfraAgentData
  • POST /api/v1/chat
    29.78s
  • agent.run
    28.90s
  • call_llm
    731ms
  • tool searchProducts
    3.62s
  • GET /search
    3.60s
  • SELECT products …
    3.41s
  • call_llm
    2.31s
  • tool checkInventory
    3.64s
  • gRPC Inventory.Check
    3.60s
  • GET sku:SONY-SEL70200GM2
    3.58s
  • call_llm
    1.89s
ErrorToolExecutionErrorcheckInventory · 23:05:20.548
{ "sku": "SONY-SEL70200GM2", "quantity": 1 }

The agent looks like it failed. It didn't. redis timed out three levels below the tool call. Without the infrastructure spans in the same trace, that line is invisible.

Architecture

A database built foragent-driven debugging

Explore the technology

Why Oodle is fast

  • Maximum concurrency: thousands of serverless invocations per query.

Why Oodle costs less

  • Object storage: 20× cheaper than disk based storage.

How Oodle scales on demand

  • Serverless compute scales up for each query, then shuts down when the work is done.

Architecture payoff

Keep everything. Query anything.
Pay 80% less.

Object-storage economics and per-query compute, without sampling, shorter retention, or fewer users.

Illustrative monthly cost

Oodle

80% lower

$2.35K

Grafana Labs

$6K

Datadog

$10.8K

Same representative workload. Expand below for assumptions and line items.

See detailed breakdown: Datadog and Grafana Labs
$15k$10k$5k$0

Common Parameters

0 GB1500 GB
0 GB1000 GB
05M
30d365d

Detailed Comparison

4.6x cheaper with Oodle
DatadogView official pricing$10,756/mo
ItemRateCost/mo
Logs1.6KB/event$2.50/M events$5,033
Traces150GB + 1M spans free/APM host · 30d indexed retention · 1M spans ≈ 10GB$0.10/GB ingested + $2.50/M indexed$438
Metrics$5-$1/100 custom metrics (tiered)$5,000
Hosts$15/infra + $31/APM host$230
Containers5 free/host$0.001/container-hr$55
Total$10,756
OodleView plans$2,350/mo
ItemRateCost/mo
Data Ingested (Logs + Traces)$0.30/GB$1,350
Metrics$2.00/1K ATS/hr/mo$1,000
Total$2,350

30 days retention included at no extra charge. Increase retention above 30d to see additional storage cost.

*Metrics: 1 sample per time series every 60s (ATS = active time series / hour). Retention: 30 days included; additional storage billed at $0.001/GB-month. Usage rates shown; see plans above for minimum commitments.

Calculate your savings

Migration

One platform to replaceDatadog, Grafana, Elastic, and Langfuse

Import dashboards, alerts, evals, and prompts. Keep your existing instrumentation.

100% open standards

PromQL and Grafana-compatible metrics · TraceQL for traces · Lucene-compatible logs · OpenTelemetry native

Elastic

< 1 day

effort over 1 week

Grafana

< 1 day

effort over 1 week

Datadog

1 week

effort over 3 weeks

Langfuse

< 1 day

effort over 1 week

Bedrock Data
“It took an hour to onboard the data.Getting us up and running took4 or 5 hours end-to-end.”
Olaf Stein

Olaf Stein

Deployment models

Your data, your rules.

SaaS, your bucket, or your cloud.

SOC 2 Type II · GDPR compliant · ISO 27001 · HIPAA ready · Built-in PII redaction

SaaS

Fully managed

We run it. Live in minutes.

BYOB

Bring your own bucket

We process. You keep the storage.

BYOC

Bring your own cloud

Full stack in your VPC. Data stays put.

Get started for free

Onboard in two commands.

# install the skill pack

# ingest metrics, logs, traces and import dashboards, alerts

FAQ

Frequently asked questions

Why does the architecture matter - isn't this just another observability tool?

The architecture is the product. Legacy observability platforms were built on disk-backed, fixed-compute infrastructure - designed for humans generating logs at human speed. When AI agents generate telemetry at machine speed (10,000 events in 3 seconds, 1,000 parallel queries per incident), that architecture either collapses under load or bills you into oblivion. Oodle separates storage from compute: telemetry lands on S3 (elastic, cheap, never needs provisioning), and queries run on serverless compute (spins up per request, scales to any concurrency). The result is full-fidelity telemetry you can actually afford to keep, and query performance that doesn't degrade as your data grows.

A single agent workflow can trigger 10,000 log events in 3 seconds. Legacy platforms bill per event. At 100 agents in parallel, that's $14,000+ per month in unexpected charges - before you've even asked a question. Disk-backed architectures also fall over when agents fire thousands of parallel investigative queries; fixed compute clusters weren't designed for that pattern. Oodle stores everything on S3 and queries run serverlessly, so cost stays flat and performance scales automatically.

Is it really 80% cheaper? How should I think about the cost model?

The difference comes from how Oodle charges: by data volume (GB ingested), not by events, spans, hosts, or seats. At 100GB logs/day, 50GB traces/day, and 500K active time series, savings compound at scale versus Datadog's per-event/per-span pricing. Verified by Lookout (80% reduction), Curefit (3×), and others. Calculate your numbers →

How long does drop-in migration actually take?

Grafana and Elastic: 2–3 hours to connect, a day or two to validate. Full migrations (the scale of Lookout - 1,000s of nodes, 2,000+ alerts, 300+ dashboards): 2–6 weeks. The process is 1-click dashboard and alert import, keep your existing agents with no code changes, run both systems in parallel, cut over when you're confident. View migration guides →

Talks + writing

How we think about observability for agents

Debugging AI agent failures at scale is hard

Enterprise MCP · Workato · 2026

Debugging AI agent failures at scale is hard

Building a logs engine faster and cheaper than ClickHouse

The Fifth Elephant · 2026

Building a logs engine faster and cheaper than ClickHouse

You Can't Sample Your Way to Reliable Agents

From the blog · Vijay Karthik

You Can't Sample Your Way to Reliable Agents