Oodle.ai Launches Agent Observability at $10/Million Traces

Oodle.ai (by Kiran and Vijay) launched an Agent Observability product priced at $10 per million spans, with sub-second P99 query latency and zero sampling. The service stores traces in S3 using a custom Parquet-like columnar format, queried via AWS Lambda.

Key details from the HN launch:

• Custom columnar storage engine built over two years for logs, metrics, and traces, now applied to LLM agent traces.
• Traces can be MBs to GBs; stored in S3 in a proprietary parquet-like format, queried serverlessly with Lambda.
• Deterministic analysis per span before LLM evals: detects tool failures, retries, loops, abnormal token usage, latency regressions, schema violations, sentiment, and other production signals.
• Pricing: $10 per million spans. Ingestion $0.30/GB, retention $0.001/GB/mo (example: 1200 GB/mo + 90 days retention = $362/mo).
• Out-of-the-box insights: Error Recovery Failures, High Duration, Low User Satisfaction, Model Cost Optimization, Caching Inefficiency, Excessive LLM Turns.
• Authors previously used Langfuse (6x more expensive). Currently processing 3M+ agent traces/day with zero sampling.

How it works: Oodle stores all traces without sampling, analyzes each span deterministically for failure signals, then optionally runs LLM-based evals on flagged traces. Queries across retained data are uniformly fast — no warm/cold tiers.

Who it's for: Engineering teams shipping AI agents in production who need affordable, full-fidelity tracing without sampling.