What Will You Build With Coding Agents That Don’t Quit? | DevDay 2026
Give a coding agent a clear outcome, a way to verify its work, operating rules, and room to delegate so it can handle recurring tasks with less supervision.
- Choose a recurring task, define what success looks like, and let the agent determine the steps.
- Add repository rules and progress notes; connect relevant tools and use separate threads or worktrees for parallel work.
- Set escalation and resource limits, and use scheduled follow-ups for tasks waiting on reviews or other events.
The CI-agent case study and feature walkthrough show how a long-running workflow is structured in practice. · Best for: Engineering teams automating CI triage, debugging, and other recurring repository work.
The speakers describe building 3D projects with coding agents, then explain how an agent handled recurring CI failures for six weeks. They cover the setup and Codex features used to run, delegate, monitor, and constrain long-running tasks.
3D projects built from broad goals0:19
- Codex used seven reference photos to build a Ferrari model, including clay and wireframe versions.
- For a Nürburgring fly-through, Codex researched references, planned the work, and used sub-agents to check track geometry.
- The render took a little over two days; Codex resumed after interruptions caused by resource problems.
A long-running agent for CI failures3:26
- The speaker contrasted prescriptive plans with giving the agent an outcome, trade-offs, and success criteria.
- A CI agent ran for six weeks, investigating flaky tests and proposing fixes without repeated prompting.
- After four days, sub-agents were assigned the slow, individual debugging work while the main agent coordinated tests and communication.
- The agent opened more than 100 pull requests per week, pushed more than 1,000 commits, and merged more than 213 fixes.
- For a difficult intermittent GPU test failure, it correlated logs, reproduced the issue on a particular node, and worked with system researchers.
- When a fix might simply delete a test, it sometimes used another agent to review the proposed solution.
Set up a long-running task12:14
- Use one thread per desired outcome; Codex can also use `/goal` to keep pursuing a task.
- Write repository guidance in `Agents.md`, including test instructions, do's, off-limits actions, and progress notes.
- Compaction is built in to help retain context over long tasks.
- Use a remote environment or Codex Cloud to keep work running without a laptop staying on.
Delegate, steer, and follow up17:55
- Connect plugins such as GitHub, Linear, and Slack so the agent can find relevant issues and discussions.
- Split independent work into child threads; use a separate worktree for each thread to isolate code changes.
- Use `/side` to ask questions in a fork of the current context without altering the main thread.
- Use Ultra for a particularly hard problem; use separate threads for work that can be split in parallel.
- Set a heartbeat to revisit a task periodically, such as checking for pull-request feedback; scheduled tasks can also start recurring summaries.
Cost and getting started23:30
- The speaker estimated the six-week CI run at about $32 per hour, then said a newer model reduced that to $16 per hour.
- Start with time-consuming work such as backlog tasks, log monitoring, or CI failures.
- Give the agent a way to test its work and recognize completion; specify escalation rules and limits on resource use.
Step by step
- Pick a recurring task that takes time away from higher-priority work.
- Define the desired outcome, trade-offs, and how the agent can verify success.
- Add repository instructions covering test commands, constraints, escalation, and progress notes.
- Run the agent in a suitable local or remote environment; connect relevant project tools.
- Delegate independent tasks to child threads and worktrees; use Ultra for unusually difficult problems.
- Set follow-up checks for external blockers and specify when the agent should stop or request help.
- Review outcomes and adjust instructions and resource limits as needed.
Tools & resources mentioned
- Codex — coding agent used for long-running development and CI work.
- Astra — model used with Codex for the Ferrari 3D project.
- Blender — software used to render the racetrack fly-through.
- Codex Cloud — managed virtual machine environment for agent tasks.
- Agents.md — repository guidance file for agent rules and project context.
- GitHub, Linear, and Slack plugins — sources for issues, discussions, pull requests, and coordination.
- Worktrees — separate code checkouts for parallel threads.
- Ultra — mode for giving a difficult problem more reasoning and sub-agent support.
- Heartbeat and scheduled tasks — mechanisms for follow-ups and recurring work.
Notes are AI-generated from the video's transcript and may miss nuance.