A multi-agent command center that turns session traces into company knowledge

Problem

If you're running agents across more than a handful of projects, you've probably hit two walls at once: orchestrating concurrent agent work without drowning in windows and tabs, and watching a technique you fought hard to get right on one project stay siloed there instead of helping the next one. The Focus AI runs into this at scale — roughly six or seven active clients and around 150 repos in its GitHub org — but the pattern bites anyone whose work spans multiple codebases. Every project keeps re-solving problems its neighbors already solved, and every solved problem evaporates into a session log no one revisits.

Approach

Will Schenk turned his laptop into a keyboard-driven agent command center and layered two systems on top of it:

• Workstation. The Aerospace window manager with key bindings jumps between terminal, mail, Slack, and browser. All agents run in Cmux — projects on the left, per-project tabs as parallel work threads — switched with command-1/2/3. He no longer reads much code; he watches what agents do and judges whether a PR is functional and tested. He'd previously moved from Emacs to Cursor to this, finding the IDE-centric agent view less useful once work became many concurrent threads.
• Remote agents over Tmux. Claude runs on a remote Linux box via Tmux — a separate Claude user, because he'd maxed out his Max plan — so work keeps running after he detaches and closes the window. He flips between agents and likes Pi for switching models (Grok, GLM, open-weight).
• A 'standards' repo. Best practices — deployment, security skills, 1Password/Infisical for env vars, preview builds — are written up once, then pointed at any older repo to bring it up to current conventions and ship the changes as a tested PR (his example: a project built months ago on a now-outdated task-tracking setup).
• 'Habitats'. Reads the session traces left by different agents (anti-gravity, Claude Code, Pi) across a project, classifies each session (plan vs. feature), summarizes the conversational beats and pivots, and extracts facts — routed either into a project's agents.md or up into the company-wide standards repo.

On self-improvement, his rule is that it only works when the task is verifiable — fuzzy creative tasks (he cited having an agent write an invoice cover letter) don't get better that way. And he's pulling back from automated outreach because recipients now recognize and discount it.

Results

Will frames the impact qualitatively rather than with metrics. Older projects get brought up to current conventions through tested PRs instead of manual refactors, and per-project learnings get consolidated into reusable, company-wide skills — his example being a hard-won technique for getting Claude to call nano-banana to generate on-brand invoices, which became a shared skill every project and customer can use, not just the one where it was first solved. He also points to where improvement actually comes from: the most frustrating traces (where you're swearing or hitting the same bug repeatedly) are the most useful — a one-line-on-page-two invoice bug grew the PDF skill from a simple markdown-to-PDF pass into a whole build system behind a single /report command.