Tony's take: Grok Bot leads.
My leading Shepherd is Grok Bot. I put it ahead of OpenAI Codex and Dots; Dots is not convincing me in the same way. That is my current take, recorded on 7 October, not a universal verdict.
The higher Grok Bot marker records this personal preference. It does not claim new evidence of better controls or operational safety. The 1 October source reading and trial caveats remain, and Dots' source-led position is unchanged.
5 October: Muse joins the watch.
Tony's Muse article asks a practical question: could it help prepare a meeting, bring back the important exception and draft the follow-up? These are proposed uses, not workflows he has tested.
- Sourced evidence · Meta MuseMeta's design account documents ongoing goals, side chats, editable memory and activity. This is the Muse personal-agent product, not a ranking of the Spark model or Muse Code.
- Bounded actionSentinel checks scoped permissions, including grants that last beyond a single action. Side chats and subagents share the user's computer. An activity view does not establish complete receipts or recovery in our own work.
- Editorial inference · provisional placementMuse enters the connected-context side, just below the Shepherd zone. Useful two-way Muse voice is unverified in the sources checked. WhatsApp messaging, Meta AI voice and Voice Transcribe are not interchangeable evidence. The open voice dot and first-watch star make those limits visible.
- Availability and costMeta's September business announcement names the US and Canada. The official download listing describes Mac and mobile clients; this is not a UK account-entitlement receipt. Subscription help lists a free allowance, Power at $20/month and Maximum at $100/month, with eligibility limits. No GBP price or local-only processing claim is made.
- Tony's stated view · unchanged pickHis article proposes a small, read-only meeting brief before expanding responsibility. It explicitly says he has not run those workflows. Grok Bot remains his personal October pick; adding Muse does not invent a new verdict.
- Watch nextCheck two-way voice in the same ongoing relationship, grant revocation, stopping background/delegated work, source-linked receipts and recovery on one comparable bounded task. Privacy and account suitability matter before company information is connected. No trial was run for this addition.
A new personal pick. A clearer idea of governed.
The useful change is not another clever model. It is a closer connection between thinking aloud, finding the right work and asking for bounded action. October opens with a new personal favourite, and a field that is still learning.
What do we mean by governed?
A governed agent pursues named work within named authority, keeping human decisions, evidence and recovery intact. More autonomy is not automatically better governance.
- Name the work. An owner, outcome, scope and finish line.
- Bound access. Least privilege and enforced limits, not only instructions a model tries to follow.
- Ask at human gates. Approval for consequential actions and the right audience.
- Show the receipt. Inspect intent, permission, execution and result. A transcript or “done” is not complete proof.
- Stop and recover. Know what happens to the main task, delegated work and schedules, and what cannot be undone.
Role is separate from proof. A goal-led worker pursues a defined outcome; a continuing coordinator carries several responsibilities. Both can be useful. Voice and continuity make a Shepherd relationship; they do not replace the controls.
Three evidence states: controls documented → bounded trial checked → sustained conduct observed. Our new October comparison is at the first stage. Strong source confidence is not a measured safety score, and no candidate is certified a steward.
Evidence: what is documented · Inference: our reading · Tony's view: his dated personal call. A missed approval or ineffective stop cannot be averaged away by good voice or memory.
- Tony's stated view · OctoberGrok Bot is Tony's winner this month. This is his dated personal call, not an objective award. Codex's earlier place in his working life remains in the archive.
- Sourced evidence · Grok BotThe documented offering combines live voice, persistent named Bots, a cloud computer and desktop/mobile access. Access expanded in August; October is its first observation here, not its launch date.
- The boundary mattersEnterprise action recording is optional and off by default, and model-based review does not gate every side effect. Enterprise network policies are not universal self-serve controls. An activity transcript is not automatically a complete audit receipt.
- Sourced evidence · OpenAI DotsDots adds a continuing cloud agent alongside Codex, with voice calls, relevant cross-channel memory and delegated Work/Codex tasks. Our first placement is provisional; Tony has not made a personal call on it.
- Dots: access and stop linesRollout depends on plan and region: personal Pro access currently excludes the UK, EEA and Switzerland; Business Premium and Enterprise roll out worldwide, with Enterprise admin enablement. Mobile needs the supporting app update. Pause stops the main task, not delegated tasks or schedules; custom rules can make mistakes and stopping cannot undo completed actions. Activity is inspectable, but complete receipts and recovery remain untested here.
- Editorial inference · assessment recalibrationDots moves further right for its continuing responsibilities and coordination. Codex moves left within connected work as a goal-led worker with thread-scoped continuity; it is not merely a one-shot coder. Neither moves upward in this recalibration. Codex retains its execution-boundary credit. This clarifies our reading, not a product regression; July–September stay unchanged.
- Editorial inference · cautious movementClaude's unified conversation rollout moves it a little right and up; two-way voice is now documented across devices. Cursor Projects earns a continuity move. Notion's shared skills and confirmation gates earn a small contextual/action move. None is a measured score.
- Open-source watch · voice is changinggoose adds live desktop voice and recipe consent; it moves modestly up while staying early-confidence. Rowboat Spaces brings shared teamwork with personal context kept on each user's machine; its position is held. No new runtime was installed or qualified.
- What stays unproven?Microsoft's Autopilot is private preview, not universal access. Codex cloud work and steering do not imply one merged history. Perplexity Computer changes are not automatically credited to Comet. Other placements hold with their earlier evidence dates visible.
- Cost and availabilityCheck the plan, device, region and beta status before choosing. Grok Bot access depends on an eligible subscription. Compare the cost of a completed, reviewed task, including supervision and recovery; this pass is not a price benchmark.
- What we watch nextRun the same low-risk test: speak a goal, recover the correct thread, accept a correction, prepare an action, stop for approval, show the result and recover after interruption. Any missed approval, lost context or weak receipt can change the next placement. The star means first observation, never quality.
The opening review was checked on 1 October 2026; Muse was added after a separate source check on 5 October. Tony's Grok Bot and Dots views were updated on 7 October without a new vendor-capability review. Other source dates and placements are held. Evidence, interpretation and Tony's personal call stay separate; no tool is hands-on qualified by this update.