AI & Automation Briefing - July 18, 2026
DoorDash Launches Agent-Native CLI, Signaling a Broader Shift in How Services Are Built
DoorDash has released a limited beta of dd-cli, a command-line tool that lets AI agents search stores, find deals, and complete food orders programmatically. The tool is available to U.S. and Canadian macOS developers via waitlist. Rather than replacing a consumer app, dd-cli exposes DoorDash's ordering platform as a building block for developer-built tools and agentic workflows. Developers can combine it with other capabilities to create custom ordering experiences without touching the DoorDash app directly. DoorDash already surfaces its service through ChatGPT and Claude, and previously experimented with iMessage ordering. The dd-cli launch is a concrete example of a growing pattern: consumer platforms redesigning their interfaces to serve AI agents rather than human users, expanding the API surface area available to anyone building agentic systems.
SEED Framework Targets Mid-Task Failures in Multi-Step AI Agents
A new training method called SEED (Self-Evolving On-Policy Distillation) addresses a core weakness in reinforcement learning for agentic LLMs: trajectory-level rewards only signal whether a task succeeded or failed at the end, giving the model no guidance on which intermediate steps went wrong. SEED works by reviewing completed trajectories in hindsight, identifying better decisions at each step, and distilling that guidance back into the model during ongoing training. The approach is designed for multi-turn, tool-using agents where long sequences of decisions compound errors before any reward signal appears. For teams building agentic workflows with tools like n8n and Claude, this research explains a structural reason why agents degrade mid-task and points toward training approaches that produce more reliable step-by-step execution.
SearchOS-V1 Gives Multi-Agent Search Systems Persistent Shared State to Break Repetitive Loops
Researchers released SearchOS-V1, a framework for multi-agent collaboration on open-domain information retrieval tasks. The core problem it addresses is a common failure mode where agents repeat the same searches because no shared record of progress exists across the agent network. SearchOS converts implicit search state into explicit, persistent state that all agents in a workflow can read and update. For builders running multi-step agent chains in tools like n8n or Claude, this maps directly to the challenge of keeping sequential or parallel agents synchronized without redundant work. The paper provides a concrete architecture for thinking about state persistence as a design requirement rather than an afterthought in agentic pipelines.