Home/Docs/Current Setup

How Adam's setup works today

Reviewed October 30, 2026. This is a dated description of one personal workspace, not a promise of identical capabilities on your account or device.

The short explanation

ā€œI run a human-supervised personal AI workspace. OpenClaw provides an always-on conversational assistant, connected tools, memory, and scheduled automations. Paperclip coordinates persistent specialist assignments and their work history. Mission Control is a private dashboard for agent activity, decisions, and system health. Small requests can be handled directly; larger work can go to specialists, with human review for consequential actions.ā€

What Adam observed in the October 30 checks

  • Direct assistant work: OpenClaw is the ongoing conversational and tool interface. Not every request is delegated.
  • Paperclip execution: a real specialist launched and completed a bounded health check after the October 30 runtime repair.
  • A useful handoff: a separate communications specialist produced architecture and workflow drafts used in this learning-site update.
  • Proactive teaching pilot: deterministic metadata checks identify relevant changes and dispatch bounded specialist review batches. The initial email and grading-readiness batches completed: the email review avoided redundant replies, and the grading review reconciled existing work instead of reposting grades. These are finite verified runs, not a long-term reliability guarantee.
  • Usage controls: unchanged checks do not need model calls. The new teaching dispatcher permits at most three batches per day, with separate category caps and one active batch per workflow. These are application controls, not a guarantee of a particular provider allowance or bill.
  • Review boundaries: student and grader replies remain private review drafts; count monitoring is not automatic scoring, grade publication, or message sending.

This is a bounded pilot, not a claim of long-term unattended reliability. Background checks remain dependent on source access, signed-in applications, provider limits, and the host being available. Cached email is not proof of server synchronization.

These are author-observed checks, not independently reproduced installations. See review scope and verification limits for the release, scenarios, and limits.

October 31 routing update

Riley's communication drafting now uses Claude Opus through the existing Claude Code subscription route. A real Paperclip assignment completed with an editorial review of the 25 published blog posts and 41 learning-site pages supplied in its public source packet. The coordinating assistant inspected the resulting artifact and run metadata. This is one verified handoff, not a claim of long-run reliability.

The role remains on demand with a maximum of four daily runs and one concurrent run; periodic role heartbeats remain disabled. Normal coding and routine observation routes were not switched. See model routing and GitHub/Gitea workflow choices.

Roles are not activity

There are 16 configured roles. Most are on-demand, and a few demonstrated workflows are more valuable than an impressive-looking roster. Portfolio ownership, model choice, and operating-system isolation are different concepts.

Current roles explains responsibilities. Architecture explains the layers. Proactive workflows explains the general pattern.

What this site does not expose

The public learning hub is separate from the private operations dashboard. It does not provide access to personal mail, student work, client systems, internal infrastructure, authentication material, or private task records.

What changed from the earlier material

The earlier guides described a Claude-only setup, older agent names, and configuration examples that should not be treated as current product commands. The refreshed written guides replace that model with current responsibilities, official setup references, explicit permissions, and bounded automation. The unreviewed short videos have been withdrawn from the current site. Use the written guides; no video walkthrough is currently offered.