Week 8 of 10

Background Agents

Core ≈ 3 h 30 min4 videos4 criteria

Focusasynchronous, cloud-delegated agents; managing fleets of parallel agents; issue-to-PR pipelines and triggers from Slack, Linear, and GitHub.

You buildAn issue-to-PR flow with isolation, budgets, checkpoints, and retries

Core material≈ 3 h 30 min

  1. 1
    The products
    Codex cloud
    OpenAI

    ; Anthropic, Claude Code on the web; Cursor, Cloud Agents; GitHub, Managing issues and pull requests with the Copilot coding agent (≈ 20 min each; compare their trigger, isolation, and approval models).

  2. 2
    Orchestration case study
    Open-sourcing Symphony
    OpenAI

    And the Symphony specification (≈ 60 min) — issue-to-PR orchestration as a specification.

  3. 3
    Video37 min
    Cloudflare's AI Engineering Stack: Assisted to Delegated
    Rajesh Bhatia of Cloudflare

    How Cloudflare moved from assisted coding to delegated agents on its own platform.

  4. 4
    Durable execution15 min
    Scaling AI Agents Without Breaking Reliability
    Preeti Somal of Temporal

    Persistent state, orchestration, visibility, automatic recovery, human interaction, and parallel work.

  5. 5
    Fleet operations9 min
    I Run a Fleet of AI Agents Across Three Machines: Here's What Broke
    Kyle Jaejun Lee

    Attention limits, durable context, blocking approvals, routing, conflicting work, and multi-machine recovery.

Tools and references

Additional video track

  • Build your own harness1 h 52 min
    Claude Agent SDK: Full Workshop
    Thariq Shihipar of Anthropic

    The SDK that Claude Code itself runs on, for building a custom background agent.

  • Where agents are going22 min
    Building the future of agents with Claude
    Anthropic

    .

  • Infrastructure economics1 h 23 min; the relevant chapters are long-running agents at `05:32`, the inference stack at `15:12`, throughput versus latency at `20:03`, and transformer hardware at `33:19`
    Why AI Is About to Get 1000x Cheaper
    Neil Movva

    . Treat “1000x” as an ambition, not a demonstrated forecast.

Build

Build an issue-to-pull-request background flow. Each job gets an isolated worktree or container, a strict budget, a checkpoint, deterministic validation, and a resumable state. Add retries with bounded backoff and prevent two agents from editing the same work item simultaneously.

Run three jobs in parallel: a small feature, a bug fix, and a documentation task. Interrupt one midway and demonstrate that it can resume or fail cleanly. Agents may open pull requests; only a human may merge.

Done when

0/4