# Personal Agents Named M I've been thinking about what a **personal agent** actually is, stripped of the marketing: software that converts expressed intent into completed real-world action. The user states an objective conversationally, the procedural middle — searching, comparing, booking, purchasing, negotiating, scheduling, following up — is surrendered to the system, and a completed result returns. It answers the question "what should exist in the world" rather than "what is the answer." The scarce commodity in the arrangement is **agency**: effective time and executive capacity, not information. Here is the curiosity: people have lived like this before. From 2015 to January 2018, **Facebook M** — a personal agent inside Messenger — served roughly 10,000 Bay Area users who could hand it objectives instead of questions: make the reservation, buy the thing, plan the trip, negotiate the bill, deal with the cable company. Contemporary reporting called access a fantastic perk, and [[wiki/Meta|Meta]] itself described users as passing off chores and "extending their bandwidth." For those users, a request disappeared into Messenger and came back accomplished. The architecture behind the curtain is the interesting part. Contemporary reporting put M's automation at roughly 30%, with **human trainers** completing the remaining ~70% — while Meta recorded the trainers' steps as training material for progressively more capable automation. Whether silicon or carbon performed the intermediate steps barely mattered to the people using it. What M demonstrated was demand for the agent abstraction itself. What it could not demonstrate was a way to supply it, because **human compute does not scale** — and M was free to its users. So those 10,000 people held something stranger than early access to a chatbot: a temporary **asymmetry of agency**. While everyone around them paid full price in time and attention for bookings, negotiations, searches, and bureaucratic recovery, they had externalized entire classes of tedious objectives. The deeper privilege was behavioral — two years of learning which objectives can be handed off, how much context to supply, and how to reorganize a day around the assumption that intent converts into action. That is **agent-management literacy**, and they were learning it a decade before everyone else. What changed between 2018 and 2026 is the supply side. **Cheap machine cognition** — planning models capable of multi-step tool use — replaced most of the human execution layer. **Persistent secure computation** (agent virtual machines that outlive the application session), **browser-use competence**, **memory**, and a **payment and approval substrate** completed the stack. Meta's 2026 [[wiki/Meta Muse|Muse]] is the scaled implementation of the same product ontology M tested — express intent, surrender the middle, receive the result — with the human-to-machine ratio inverted. What did not change is the boundary where machine execution still meets resistance: the telephone. In September 2026 Reuters reported, from internal Meta posts, that **human contractors** had been quietly placing Muse's phone calls to businesses — pushing completion into the 95–98% range against a lower AI-only rate. Employees raised disclosure and privacy concerns — users were not told a human was on the line — and a Superintelligence Labs vice president called it "a miss" and rolled the feature back. The pattern is M's pattern compressed: **human labor now concentrates at stubborn boundary conditions** — voice interfaces where counterparties will not engage synthetic callers — rather than across the whole service. The ideal setup, as I've worked it out: **delegate outcomes, not procedures**. The formulation that works is a completed state plus constraints — what should exist in the world, by when, within what budget, under which hard limits. Vague objectives are acceptable, since resolving them is the agent's job, but constraints must be explicit, because the agent optimizes for completion over cost unless told otherwise. I supply context once and let memory carry it; the setup cost of a delegation falls with every repeated task. **Keep approvals on irreversible actions**: purchases, sends, bookings, and anything touching money or other people gate on my approval, while reversible work — research, comparison, drafting, monitoring — runs unattended. The [[wiki/Approval Architecture|approval architecture]] is the actual control surface of the product, and I configure it before delegating anything consequential. I stay aware of where the human layer sits — agent-placed voice calls may currently be handled by human contractors without disclosure, so I avoid delegating sensitive negotiations through agent phone-calling until disclosure practices are settled, and I treat call transcripts the way I treat any third-party account: verify before acting. **Verify, don't re-do**: the characteristic failure mode is re-performing the delegated work out of distrust, which surrenders the entire economic gain. I spot-check results against independent sources — confirm the reservation exists, the price matches — rather than shadowing the process. And the durable skill is the one the M enclave learned first: **learn the delegable classes** — which objectives externalize cleanly, how much context a task needs, how to run several delegated tasks in parallel. The advantage compounds. Each delegation teaches thinking at the level of objectives instead of procedures. ## Related topics - [[wiki/Facebook M|Facebook M]] - the 2015-2018 Messenger enclave: 10,000 users, ~30% automation, human trainers as the execution layer - [[wiki/Personal AI Agent|Personal AI Agent]] - the general category: intent in, completed real-world action out - [[wiki/Meta Muse|Meta Muse]] - the 2026 consumer-scale implementation of the same product ontology - [[wiki/Meta|Meta]] - operator of both Facebook M and Muse; twice inserted human labor behind the agent interface - [[wiki/Artificial Intelligence|Artificial Intelligence]] - the substrate: planning models, tool use, and memory replacing M's human execution layer - [[wiki/Agent-Management Literacy|Agent-Management Literacy]] - the durable skill the enclave learned first: which objectives hand off cleanly - [[wiki/Approval Architecture|Approval Architecture]] - approvals on irreversible actions, autonomy on reversible work - [[wiki/AI Counterpart|AI Counterpart]] - the agent as counterpart rather than tool - [[wiki/Agentic Empowerment|Agentic Empowerment]] - the empowerment thesis this history supports - [[daily/20260927 - The Age of Agentic Empowerment|The Age of Agentic Empowerment]] - companion entry: agents misclassified as productivity tools, leverage over allowance