A landscape map of browser automation tools for AI coding agents in mid-2026 — Playwright CLI, Vercel's agent-browser, Stagehand, Browserbase, ReportPortal, and where Claude for Chrome fits.

A competitive-landscape report of four GitHub repos across agent context engineering, agent memory, and data cataloging — where they overlap, where the gaps are.

A durable message bus that lets independent Claude Code instances in different tmux windows coordinate and hand work to each other.

Coding agents work better when they write to disk, but those files end up scattered and lost. scratch organizes an agent's working memory into durable, structured, shareable scratchpads you can browse.

Add OpenTelemetry traces, metrics, and logs to the pi coding agent — one trace tree per turn, exported to Aspire, Grafana, or any OTLP backend.

An observability dashboard for pi.dev — sessions, todos, subagents, and pinned context in one place.

Expose Hangfire jobs and maintenance tools as MCP capabilities, so AI agents can run jobs and operate the queue. Aspire integration is a bonus.