# What Will You Build With Coding Agents That Don’t Quit? | DevDay 2026

## Executive summary

The presentation details how advanced AI agents (Codex, Astra, GPT-6) can automate complex, long-running engineering tasks, extending agent capabilities from simple code generation to sustained, multi-week maintenance of Continuous Integration (CI) systems. Key advancements include the ability to manage context over long horizons, parallelize work using child threads, and integrate with external tools (GitHub, Slack) to autonomously debug flaky tests, manage pull requests, and coordinate fixes, significantly reducing manual engineering overhead.

## Key takeaways

- Sustained Automation for CI/CD: An agent can autonomously maintain CI health over extended periods (e.g., six weeks), running millions of tests, opening over 100 PRs per week, and merging over 200 fixes, even handling complex issues like Heisen bugs and non-deterministic floating-point errors. (See 10:30)
- Advanced Agent Control Flow: New features allow engineers to manage agent behavior by using child threads for parallel tasks, the `/goal` command for relentless task pursuit, and the `/side` command for querying without disrupting the main workflow. (See 13:20)
- Context Management and Scale: The introduction of 'Compaction' allows models to maintain context over extremely long-horizon tasks (months) without degradation. Remote environments and Codex Cloud enable continuous work across different machines and surfaces. (See 11:30)

## Technical details

- Agent Architecture & Workflow: Agents can be given a high-level goal and trade-offs, allowing them to figure out the steps autonomously, rather than requiring prescriptive, step-by-step plans. (See 8:20)
- Parallel Processing: Using child threads and worktrees allows agents to work on multiple, non-sequential features (e.g., a 3D model and a website) simultaneously without causing merge conflicts. (See 14:50)
- Debugging Hard Problems: Ultra Mode provides deep reasoning time, allowing the agent to spin up sub-agents to investigate complex, single-point failures (like the GPU floating-point bug) that require sustained, focused effort. (See 15:40)
- Integration and Monitoring: Plugins (e.g., GitHub, Linear) connect the agent to external tools, allowing it to search issues, discuss PRs, and coordinate fixes. Heartbeat and scheduled tasks automate follow-ups and periodic summaries. (See 14:00)

## Practical implications

- Identify repetitive, time-consuming tasks (e.g., manual CI debugging, backlog triage, routine testing) that can be converted into agent-driven loops.
- Implement a structured 'Agents.md' file to guide the agent's scope, rules, and limitations.
- Utilize child threads for tasks that can be broken down into parallel, independent components to accelerate overall progress.
- For critical, single-point failures, switch to Ultra Mode to dedicate maximum reasoning power to the problem.

## Topics

AI Agents, Continuous Integration (CI), Software Development Lifecycle (SDLC), Automation, Generative AI, Codex, Astra, GPT-6 Sol, Codex Cloud, Agents.md

Source: https://www.youtube.com/watch?v=U1IVZy9FhVQ
