# Building the Backbone of AI Agents: Telemetry, Open Source, and the Future of Developer Infrastructure with Brian Douglas

Page: https://stenobird.com/podcast/screaming-in-the-cloud/building-the-backbone-of-ai-agents-telemetry-open-source-and-the-future-of-developer-infrastructure-with-brian-douglas
Text version: https://stenobird.com/podcast/screaming-in-the-cloud/building-the-backbone-of-ai-agents-telemetry-open-source-and-the-future-of-developer-infrastructure-with-brian-douglas.md
Podcast: [Screaming in the Cloud](https://stenobird.com/podcast/screaming-in-the-cloud)
Published: 2026-04-30T10:00:00+00:00
Episode link: https://share.transistor.fm/s/41391d7a
Audio file: https://dts.podtrac.com/redirect.mp3/media.transistor.fm/41391d7a/34028eaa.mp3
Processing state: processed
JSON: https://stenobird.com/v1/public/podcasts/screaming-in-the-cloud/episodes/building-the-backbone-of-ai-agents-telemetry-open-source-and-the-future-of-developer-infrastructure-with-brian-douglas
Duration seconds: 1757

## Resource

AI agents are advancing rapidly, but the underlying infrastructure for observability and execution is lagging. Brian Douglas explains how Paper Compute is building the necessary telemetry and sandboxed runtimes to make agentic workflows reliable and secure.

## Highlights
- Main idea: Agentic workflows require specialized telemetry beyond standard OpenTelemetry to track tokens, duration, and prompt context
- Practical takeaway: Using sandboxed runtimes like Stereos (built on NixOS) can prevent agents from accessing unauthorized local data or system resources
- Failure mode: Relying on default LLM permissions can lead to unexpected data leaks, such as agents accessing personal calendars via connected accounts
- Main idea: Open source serves as a critical trust signal for infrastructure tools, allowing developers to audit policies and controls
- Practical takeaway: Implementing Merkle DAGs for agent sessions enables efficient session hashing, searching, and skill generation

## Topics

AI Agents, Developer Infrastructure, Telemetry, Open Source, Observability, Sandboxing, NixOS, Distributed Systems

## Chapters
- 1:00 — Paper Compute's Mission: Brian introduces Paper Compute and its focus on building distributed systems and infrastructure specifically for AI agents.
- 3:10 — Telemetry for Agents: A deep dive into tracking prompts, human context, and the limitations of current observability tools like OpenTelemetry for LLMs.
- 5:15 — Observability and Data Interpretation: The importance of seeing raw data from agent executions to understand performance and reliability.
- 7:15 — Security and Agent Sandboxing: Discussing the risks of agentic access to sensitive data and the need for controlled, isolated environments.
- 9:35 — The Tapes Product and Session Hashing: How using Merkle DAGs allows for searchable, hashable agent sessions and the development of reusable agent skills.
- 16:15 — The Cost of Agentic Compute: The high RAM and EC2 costs associated with running intensive agentic tools like Claude Code.
- 20:35 — Open Source Strategy: Why building in the open is a strategic advantage for gaining developer trust and establishing industry standards.

## Actions

- request_transcript: `POST https://stenobird.com/v1/public/podcasts/screaming-in-the-cloud/episodes/building-the-backbone-of-ai-agents-telemetry-open-source-and-the-future-of-developer-infrastructure-with-brian-douglas/transcription-requests` — Idempotently request low-priority transcript generation for this episode.
- read_markdown: `GET https://stenobird.com/podcast/screaming-in-the-cloud/building-the-backbone-of-ai-agents-telemetry-open-source-and-the-future-of-developer-infrastructure-with-brian-douglas.md` — Read the agent-friendly Markdown representation of this episode resource.

A page view does not enqueue transcription. Agents should invoke `request_transcript` explicitly when they need this episode processed.

## Transcript

Full transcripts are not published on public pages unless there is a clear rights basis.