AI Agents

AgentMail launches a signup flow designed for AI agents, requiring a curl request followed by a human OTP claim. This addresses the friction of agents needing email inboxes for verification without human browsers. The model assumes agents can handle the initial handshake but requires human intervention for final trust establishment, creating a hybrid authentication model for machine users on human-centric platforms.

Runtime provides sandboxed coding agent infrastructure for teams, allowing non-engineers to ship via Claude Code or Codex safely. Engineering defines context once—system instructions, skills, integrations—and Runtime snapshots the full environment including Docker Compose, Kafka, and Redis. This prevents works on my machine issues and secrets leaks, enabling broader team access to agent coding capabilities without risking production stability or credential exposure.

A practitioner reports shifting workflows to local LLMs, documenting Codex errors into reusable skills for Qwen3.6. Tasks include devops on VPS, PDF conversion, and browser testing. The user interacts with the OS via natural language commands for installation and configuration. While anecdotal, the workflow highlights the value of persisting agent failure modes into skill libraries to improve local model performance on repetitive technical tasks.

Simon Willison announces the first release of Datasette Agent, integrating his LLM library directly with Datasette instances. The system provides a conversational interface for data querying, extensible via plugins like datasette-agent-charts for visualization. The live demo runs on Gemini 3.1 Flash-Lite, emphasizing cost-effective inference for data exploration. This marks a transition from static data dashboards to interactive, natural language investigation, grounded in SQL execution rather than vague summaries.

Latent Space interviews Daytona's CEO on their agent infrastructure, reporting 850K daily runs and 74% month-over-month growth. The platform focuses on bare metal sandboxes rather than containers to reduce overhead for agent execution. They discuss RL evals and a new Agent Cloud designed to handle this scale. The metrics suggest significant traction in providing compute environments specifically optimized for agentic workflows rather than traditional CI/CD pipelines.

This release introduces a permission system for agent tool access, allowing tool availability to be attached to specific required_permission flags. Background agent tools now require the datasette-agent-background permission, separating interactive queries from automated tasks. This granularity is critical for production safety, ensuring agents cannot invoke privileged operations without explicit authorization scopes defined in the Datasette configuration.

Version 0.1a3 improves observability and UX with View SQL query buttons for both tables and collapsed tool calls. It suppresses empty reasoning chunks and handles truncated responses gracefully, ensuring tables display even if results are cut off. These changes prioritize transparency, allowing developers to audit the exact SQL generated by the agent rather than trusting the natural language summary of the action.

The charts plugin adds sequential color schemes for magnitude and categorical schemes for text values, plus interactive tooltips. Crucially, it checks execute-sql permissions before running queries to determine column names, preventing privilege escalation through visualization tools. This release demonstrates how plugin architecture can enforce safety boundaries even in auxiliary features like data visualization within an agentic workflow.

A new plugin enables Datasette Agent to run commands within Fly Sprites sandboxes. This extends the agent's capability beyond data querying to system command execution, but confines those actions within a ephemeral, isolated environment. It's an early release but signals a pattern where agent actions are routed through specific sandbox providers rather than executing directly on the host.