The Shift to Open Agentic Development
Warp is transitioning from a terminal-based coding assistant to a platform for "Open Agentic Development." In this model, agents handle the implementation work—planning, coding, testing, and opening pull requests—while humans focus on defining objectives, verifying outputs, and providing the product judgment that guides the system. Warp reports that 90% of its internal pull requests are now co-created by agents, demonstrating the viability of this approach for scaling engineering teams.
Orchestration via Oz
To manage these persistent, parallelized agents, Warp built Oz, a cloud-based control plane. Oz allows developers to deploy agents across local and cloud environments, monitor long-running workflows, and maintain context through:
- Context Compaction & Memory: Techniques to help agents stay focused during extended, complex tasks.
- Subagent Specialization: Using dedicated agents for specific tasks like code search and file analysis.
- Hybrid Execution: The ability to hand off workflows between cloud and local environments without losing state.
- Evaluation Pipelines: Using LLM-as-a-judge systems to verify agent outputs.
Efficiency Gains with GPT-5.5
Warp utilizes GPT-5.5 to handle demanding reasoning and coding tasks. The model has proven particularly effective for long-horizon engineering, where it can reason across large problem spaces. Compared to GPT-5.4, Warp observed a 30% reduction in tokens per task, which is critical for the sustainability of long-running, autonomous agent workflows. The company routes tasks to different model configurations based on difficulty, ensuring that complex reasoning receives the necessary frontier-level intelligence while optimizing for cost and efficiency.