Agents for the Inner Loop and Outer Loop
summary
Platform runs cloud agents (started March 2024) that handle both inner-loop editing and outer-loop automation: agents open PRs to fix failing CI and merge conflicts, trigger on Datadog-triggered fix PRs for every error-level log, and run dependency updates on schedule. Sandboxing uses Kubernetes, network policies and limited secrets. Simple natural-language skills hit ~75–80% accuracy; combining code + prompts raises that to ~95–99%. Teams use agent code review, recorded QA videos, and auto-approval for trivial PRs to reduce reviewer load.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:36Inner vs outer loop development
- 3:42Inner loop agents: when to use
- 5:30Prompt to PR automation
- 7:19Outer loop agents: triggers
- 9:30Cloud sandboxing for security
- 12:46Agents: skills vs code
- 14:59Custom dev experiences
- 16:26Future of auto dev loops
- 17:54Humans in agentic systems
- 18:36Self-improving agents
- 19:38Adopting agents in conservative firms
- 21:06PR automation vs human review
- 24:52Agents for test coverage
related talks
27mIntegrating Agentic AI Into Internal Developer Platforms To Minimize Errors and Downtime
31mInternal Background Agent Systems: Architecture and Rollout
31mAgents at Work: How GitLab Is Redefining AI Adoption for Engineering Teams
21mPlatforms for Humans and Machines: Engineering for the Age of Agents
19mAgents at work: plan - test - CR - deploy - repeat
15mMinion: Agent-Ready Dev Platform Behind 11% of Merged PRs
20mThe Missing Layer After Launch
16mWhat Comes After AI Coding Assistants