How Octolane scaled to 4.4M agent requests with Respan

About Octolane

Octolane set out to eliminate sales busywork. Instead of reps manually logging notes, updating fields, or remembering follow-ups, Octolane automates it all. The CRM updates itself using email, calendar, and meeting data, and even recommends next actions.

After YC W24, Octolane grew fast. In April 2025, they raised a $2.6M oversubscribed seed round to help bring the first AI-native CRM to market (Forbes). Today, Octolane manages $700M in real pipeline for 1,200 companies.

The challenge: debugging blind and judging agents by vibe

Octolane runs hundreds of LLM calls per user session, from identifying patterns in sales calls to enriching contact data and generating AI field updates. But without a clear view of prompt inputs, outputs, and token costs, the engineering team lost hours debugging every week.

Debugging was only half of it. The other half was knowing whether the product was actually getting better.

When every change is judged by feel, there's no way to tell whether a new prompt or model made the agents smarter or quietly made them worse. Octolane needed observability without slowing down shipping velocity.

Setup: LLM observability in 15 minutes

The Octolane team dropped the Respan proxy into their Next.js app through the AI gateway. In minutes:

• Every LLM call, prompt, input, response, latency, and token usage was auto-logged
• Slow calls and failures were surfaced in Slack instantly
• Prompt iterations were tracked with full history
• Costs were visible across models and endpoints

From tool calls to full traces

With Respan, Octolane replaced the vibe check with data. The team can open any agent run and see exactly what happened inside it:

• Tool calls logged alongside every model call, so each step an agent takes is on the record
• Full traces that follow a run from start to finish and show which span caused a failure
• Searchable spans with inputs, outputs, metrics, and metadata for every call

Looking across those traces, the team spots patterns in how its agents behave and turns them into improvements, instead of guessing what changed.

Inside the workspace: real-time debugging in action

To show what this looks like in practice, here's a live screenshot from Octolane's workspace inside Respan:

In this view, the team is inspecting a failed LLM call with full visibility: status codes, error messages, parameters, and traces, all in one place. Instead of manually backtracking issues, they can pinpoint bugs instantly and get back to shipping.

This kind of real-time observability has become part of Octolane's everyday workflow, powering a faster, more reliable AI product.

Changes: from vibe checks to full visibility

This removed the guesswork. Instead of re-running sessions to trace bugs, the team could resolve issues in minutes and get back to building.

Results: from thousands to 4.4M agent requests

Since integrating Respan:

• Agent requests grew from the thousands to over 4.4M
• $700M in real pipeline managed for 1,200 companies
• LLM debugging time dropped 90%
• Feature velocity doubled
• AI column and Action Mode shipped with confidence