Oodle AI Agent Observability Launch – $10 per Million Traces

Oodle AI Agent Observability Launch – $10 per Million Traces

Oodle AI launches Agent Observability with $10 / M span pricing

Takeaway: Oodle AI now offers a dedicated agent‑observability service that stores 100 % of LLM agent traces in S3‑backed columnar storage, delivers sub‑second query latency, and charges a flat $10 per million spans – a cost model that eliminates per‑query fees and makes large‑scale debugging affordable.

Why fast, cheap trace storage matters for LLM agents

Agent traces are large. Each trace contains prompts, tool‑call metadata, and model responses, often amounting to megabytes per conversation. Traditional APM tools struggle with this volume, leading to slow searches (minutes) and high storage costs.

Agents fail nondeterministically. Silent tool errors, incomplete tool‑call chains, or “I did it” messages can go unnoticed without full trace retention. Sampling is insufficient because rare failures are precisely the ones that need detection.

Oodle’s solution addresses both pain points by:

  1. Storing every span in an S3‑compatible object store, allowing retention for months or years at a predictable cost.
  2. Using columnar, serverless compute to achieve <1 s P99 query latency, turning minutes‑long investigations into instant look‑ups.

Core technical features

1. Columnar storage & serverless query engine

"Fast search over traces – Columnar storage, serverless compute – find the trace in seconds, not minutes."

Oodle rewrote parts of its storage engine (see the engineering blog linked below) to store spans in a parquet‑like format optimized for read‑heavy workloads. This design enables:

  • Vectorized scans that skip irrelevant columns, reducing I/O.
  • Elastic scaling without provisioning dedicated query nodes.

2. S3‑based, flat‑rate pricing

"$10 / million spans – Store 100 % of traces and keep your budget too."

Pricing is based on ingested bytes and retention days. A typical workload of 1.2 TB/month costs $360 for ingestion plus $2 for 90‑day retention, totaling $362/month with unlimited queries.

3. Out‑of‑the‑box AI insights

"Detect user frustration, tool call optimizations, anomalies before you write a single eval."

Built‑in analytics automatically surface:

  • Error recovery failures (unhandled exceptions, incomplete tool calls)
  • High‑duration traces (stuck loops, retry storms)
  • Low user satisfaction (negative sentiment scores)
  • Model cost inefficiencies (cheaper model alternatives)
  • Caching inefficiencies and excessive LLM turns

These insights require no custom rule configuration.

Real‑world adoption

Oodle cites several production customers handling 3 M+ agent traces per day with zero sampling. Notable case studies include:

  • Fello – Voice AI platform detecting silent failures across millions of daily interactions.
  • Cureskin, HappyPath, Wisdom AI, Fuel, Lookout, Zaggle, CureFit, Distacart, Workorb, Effective AI, different.ai, Labra, Bedrockdata – all listed on the product page as engineering teams using Oodle for observability.

"3M+ agent traces/day. Zero sampling. Catch silent failures on Voice AI agents." – Fello case study

How to get started

  1. Instrument with OpenTelemetry – Oodle consumes standard OTLP traces using the Generative AI semantic conventions; no custom SDK is required.
  2. Send spans to Oodle – point your existing exporter at the provided endpoint.
  3. Explore in the UI – query traces instantly, view transcript replay, run eval pipelines, and leverage AI‑powered insights.

A free tier is available with no credit‑card requirement, and the platform promises a 15‑minute onboarding.

Community feedback from Hacker News

  • Pricing concerns: One commenter noted their current vendor costs $0.75 / M, finding Oodle’s $10 / M “expensive.”
  • Technical curiosity: Another asked why Oodle uses a “parquet‑like” format instead of native Parquet.
  • Developer insight: Vijay Karthik (Oodle engineer) responded that building agent‑trace support forced a redesign of their storage engine, linking to the blog post How We Achieved $10/Million Agent Spans.
  • Positive reception: Users praised Oodle’s speed and “best‑in‑class MCP,” expressing excitement to evaluate failure patterns.

Further reading

  • Engineering blog: How We Achieved $10/Million Agent Spans – details on the storage architecture and cost optimizations.
  • Observability blog: You Can’t Sample Your Way to Reliable Agents – discusses why full trace retention is essential.
  • Architecture blog: How Oodle Keeps Observability Fast at Scale – explores serverless query design.

Bottom line: Oodle AI’s Agent Observability service provides a purpose‑built, low‑cost platform for storing and querying massive volumes of LLM agent traces, delivering sub‑second search and automatic failure detection without the complexity of separate APM tools.

Sources