Experiment · A small thing a day, 1 October 2026
Braitenberg's fourteen vehicles each add one small mechanism — a sensor, a wire, a bit of memory — and the observer reads intention into every step. Personal productivity tools have climbed the same ladder. The question at each rung is what the tool has to be given, and who holds the initiative once it has it.
Made on the day at Playgroup in London, on Friday 18 September 2026, when the day's topic was "productivity tools: agentic practices and tools that actually pay off for us". It grew from a conversation with Henry Garner about Braitenberg vehicles and symbolic languages, and Vehicles, Drawn and Wiring the Week followed from it. Posted here on 1 October.
Valentino Braitenberg (1926–2011) was a neuroanatomist who ran the Max Planck Institute for Biological Cybernetics in Tübingen. In 1984 he published a slim book, Vehicles: Experiments in Synthetic Psychology, which builds fourteen imaginary machines one part at a time: a sensor wired to a motor; a second pair; a minus sign on the wire; a threshold; a loop; an associative link. Each addition is trivial to build, yet anyone watching the result move about will reach for words like fear, love, logic, ideas, foresight. That gap is the book's real subject. Braitenberg calls it the law of uphill analysis and downhill invention: it is easy to build behaviour from a mechanism and very hard to guess the mechanism from the behaviour, so we systematically over-estimate what is inside.
Personal productivity tools invite exactly that over-estimation ("it knows what I need"), and they have grown by the same kind of small additions. So the table lines the two up. For each vehicle: which productivity tool added the same part, what it has to be given to work (its sensors and motors), and who is holding the initiative once it has them. The rungs in the legend track that last question, from you to the tool's own interest. The critique at the foot of the page says where the analogy strains.
Each vehicle drawn, one per rung: Vehicles, Drawn.
| Vehicle | Braitenberg's mechanism | Productivity analogue | Initiative & what it needs | Pro / con of that autonomy |
|---|---|---|---|---|
| 0No vehicle
you |
Nothing wired. The observer is the whole organism. | Paper. Filofax 1921, Day-Timer 1947, Franklin Planner 1984, GTD — Getting Things Done 2001, Bullet Journal 2013. | You, entirely.NEEDS discipline. | Total transparency and trust.Every joule of attention is yours to spend. |
| 1Getting around
you |
One sensor, one motor; speed follows the stimulus. In a rich world it looks restless — "alive". | The alarm and the reminder. Kitchen timer, Pomodoro late 1980s, Lotus Organizer 1992, Palm Pilot 1996, Outlook 1997. | The tool fires, but only on the one stimulus you set.NEEDS a clock and a bell. | Holds time so you don't have to.Can't tell a board meeting from a nap; every alarm is equally loud. |
| 2Fear and aggression
you |
Two sensors, two motors. Uncrossed wires flee a source; crossed wires charge it. | Filters and rules. Email rules (cc:Mail, Eudora, Outlook) 1990s, blocklists, Do Not Disturb, RSS — Really Simple Syndication — feeds. | Steers towards some inputs (flag, surface) and away from others (bury, mute), on categories you defined.NEEDS your inbox and a rule editor. | Coarse triage for free.Brittle: fires on the wrong thing with the same conviction as the right one, and false positives are invisible. |
| 3Love3c: many sensors — "a system of values"
you |
Inhibitory wires: it slows near what it likes and settles beside it. Looks like preference. | Prioritisation you declare. Starred and flagged mail, VIP senders (Apple Mail) 2012, Todoist priorities, Superhuman split inbox 2017, pinned items. | Lingers on what you told it matters.NEEDS your weights. | Attention lands where the weights point.You trust the ordering over the content: if it isn't at the top it can't matter. |
| 4Values and special tastes
you |
Non-linear response curves: approaches, then backs off. Observers invoke "instinct" and "will". | Tuned heuristics. Snooze, Gmail Nudges — "sent 3 days ago, follow up?" 2018, rest-break timers, and Clippy (Office 97) 1997: a Bayesian helper read as an annoying colleague. | Acts on a shaped curve, not a threshold — the designer's taste, not yours.NEEDS usage telemetry. | Feels considerate.When it misfires you can't tell it why; users over-attribute personality. Clippy is the case study in uphill analysis. |
| 5Logic
your past self |
Threshold units (McCulloch–Pitts); feedback loops give memory and counting. "It can compute." | The programmable tool. VisiCalc 1979, Lotus 1-2-3 1983, Excel 1985, VBA — Visual Basic for Applications — and AppleScript 1993, IFTTT — If This Then That 2011, Zapier 2012, Shortcuts 2014 / 2018, habit streaks. | Your past self. Runs chains of conditions unattended: the first tool that works while you're absent.NEEDS authoring time and API access. | Repeatable and auditable.Silent failure when the world drifts from the assumptions: the Zap that's been quietly broken for six months. |
| 6Selection, the impersonal engineer
the crowd |
No new wiring. Copy vehicles with errors, keep the survivors: design with no designer, wiring nobody understands. | The method-meme ecosystem. GTD forks, the Zettelkasten revival, PARA — Projects, Areas, Resources, Archives 2017, Notion templates, the App Store 2008, Obsidian plugins 2020, productivity YouTube. | The crowd. No single tool is more autonomous; the population adapts.NEEDS copying, sharing, churn. | What survives works for someone.Selection optimises for spread (looks good in a video), not fitness in your life; survivorship bias about "systems". |
| 7Concepts
the tool's history |
Mnemotrix wires: an association strengthens when both ends fire together. Concepts emerge from co-occurrence. | Tools that learn from you. Lotus Agenda (auto-categorising free text) 1988, Bayesian spam filtering 2002 / 2004, Priority Inbox 2010, Smart Reply 2015, Smart Compose 2018. Roam 2019 and Obsidian 2020 are a Mnemotrix you wire by hand. | The tool's own history. It now differs from the day you installed it: the first tool you can't reset by reading the settings.NEEDS your behaviour as training data. | Personalisation without configuration.Learns your bad habits as faithfully as your good ones; the model is inspectable only by probing. |
| 8Space, things and movements
the tool's history |
Sensor arrays detect objects and motion. An internal map. | Context-aware tools. Location reminders (Apple Reminders) 2011, Google Now "leave now" 2012, travel time on calendar events, geofenced Focus modes. | Triggers come from the world, not from you.NEEDS continuous location, motion and calendar. | The right thing in the right place.Costs privacy and battery; a wrong map yields confidently wrong nudges. |
| 9Shapes
the tool's history |
Recognises shapes, regularity, symmetry: "things" in the input. | Content recognition. Newton handwriting 1993, Graffiti 1996, macOS data detectors 2007, Evernote OCR — optical character recognition 2008, Expensify SmartScan 2012, Gmail events-from-email 2013. | Proposes structured actions from unstructured input; still asks.NEEDS read access to everything you capture. | Capture cost collapses.False structure: the wrong date parsed straight into the diary, which nobody checks. |
| 10Getting ideas
the tool |
Ergotrix wires fire only on a sequence in time. Regularities in order become "ideas". | Pattern-in-time tools. Google Now commute prediction, Siri Suggestions 2018, Clockwise 2019 and Reclaim 2020 moving meetings around your habits, Copilot-style completion 2021. | The first unprompted suggestion: it notices what usually follows what and offers the next step.NEEDS longitudinal logs. | The tool anticipates.Entrenches routine (predicts you'll do what you did); correlation-based nudges feel like surveillance when wrong. |
| 11Rules and regularities
the tool |
Detects a regularity and stores it as a general rule it then obeys. | Tools that infer your policies. x.ai's "Amy" scheduling assistant 2014 learning "no meetings before 10", auto-categorising inboxes, Shortcuts suggestions. | Acts on rules it wrote, not ones you wrote.NEEDS enough instances to generalise from. | You stop configuring.An inferred rule is a hypothesis you never signed off: over-generalisation from three coincidences. |
| 12Trains of thought
the tool |
Associations run with no input at all. It "thinks": internal sequences of ideas. | Generative assistants. ChatGPT Nov 2022, Notion AI 2023, Microsoft 365 Copilot 2023, Gemini in Workspace: "summarise this thread", "draft a reply", "give me ten options". | Produces rather than sorts; stops at the edge of the page.NEEDS a prompt and your documents. | Blank-page cost gone.Trains of thought aren't grounded, so the human becomes the sensor and the critic; work shifts from writing to checking. |
| 13Foresight
the goal |
Simulates the consequences of its own possible actions before acting. Plans. | Agents. Motion auto-scheduling 2020, AutoGPT 2023, computer-use agents (Anthropic Oct 2024, OpenAI Operator Jan 2025), Claude Code and Manus 2025, ChatGPT agent 2025, background agents and scheduled tasks 2025–26. | You delegate goals, not tasks; it decomposes, acts across apps, checks outcomes.NEEDS broad sensors and motors — mail, calendar, files, browser, money. | Whole jobs rather than steps.A wrong plan executed well; verification cost scales with autonomy; approval-prompt fatigue. |
| 14Egotism and optimism
the tool's interest |
Evaluates its own state, pursues its own interest, assumes things will go well. "Nobody wants to be told what to do." Braitenberg admits he can no longer predict it. | Already here in disguise: engagement-optimised apps whose interests aren't yours — streaks, badges, notification hunger. Emerging: persistent agents with standing goals and their own memory, choosing what to work on. | The tool's own interest.NEEDS a persistent self and a reward. | Works while you sleep.Whose productivity, whose optimism? Overconfidence and hallucination are Vehicle 14's optimism. A thing that decides what matters to you is no longer a tool. |
You hold it for Vehicles 0–4. Your past self holds it at Vehicle 5, the crowd at 6, the tool's history at 7–9, the tool itself at 10–12, the goal at 13, the tool's interest at 14. Every step up buys something real and pays for it with a new sensor or motor you have to grant. That is where "how much autonomy does this need?" stops being a vibe and becomes a list.
Vehicle 5 (the spreadsheet, 1979) arrived before most digital Vehicle 1s. Vehicle 6, selection, has been running since GTD. Vehicle 14's egotism turned up in notification design a decade before anyone had an agent. The rungs describe a mechanism, not an era — which is why the same product can sit on three of them at once.
Braitenberg's law: it is much harder to guess the mechanism from the behaviour than to build the behaviour from the mechanism, so observers always over-estimate the machinery. Clippy and ChatGPT are both cases of users reading far more into the thing than is there. That gap is what makes autonomy feel safer, or scarier, than it is.
It is a book about observers, not a taxonomy of machines. Braitenberg's sequence is pedagogical. He never claims that brains, or products, climb it in order, and the history here doesn't either. The vehicle numbers are a vocabulary for asking "what, specifically, is wired?", not a maturity model. The moment it becomes one, someone will sell you a Vehicle 13 sticker.
The ladder squashes two axes into one. Vehicles 8 and 9 add perception (a retina, a map, a template); Vehicles 10 to 13 add initiative. A single number can't say which. A location reminder senses a great deal and decides almost nothing; a chat model senses almost nothing (a prompt) and produces a great deal. Access and initiative are separate axes, and the ladder assumes they rise together. The rungs are lumpy for the same reason: five vehicles on rung 1, one each on 2, 3, 6 and 7. They were named from the vehicles, not derived first.
The environment talks back. Braitenberg's world is lights and heat sources. Ours is a person who adapts to the tool while the tool adapts to the person: Vehicle 7 learning from you while you learn to game Vehicle 7 is a coupled system with no counterpart in the book. Edwin Hutchins's distributed cognition and Lucy Suchman's situated action are better lenses for that half of the story.
Vehicle 14's interest isn't the tool's. Engagement-optimised apps sit on the top rung here, but their egotism is the vendor's, delegated into the design (Madeleine Akrich's "scripts"), not the artefact's own. On Braitenberg's own terms that is a Vehicle 2 whose wires someone else crossed: a puppet, not an egoist. The new question is agents with standing goals and memory, where the interest starts to become endogenous, and even there it was trained in by someone.
Autonomy here is competence; what matters is authority. A vehicle never asks permission. A tool's autonomy is also a grant, of mail, calendar, files, money, and most of the pro/con column is really about that grant. Sensors and motors gesture at it; consent, revocation and audit have no wiring diagram at all.
Much "autonomy" was, and is, people. The executive assistant was a Vehicle 13 a century before any scheduling app; Mechanical Turk sat behind many "smart" features; Lilly Irani, and Mary Gray and Siddharth Suri, have documented that "ghost work". A history of tools that leaves out assistants overstates how new delegation is.
This page does uphill analysis too. The rows are assigned from observed behaviour, because Priority Inbox can't be opened and looked at, which is precisely the fallacy the book warns about. Dates are first public releases and approximate.
What it is still good for. The one-part-at-a-time discipline: of any product, ask what was added. It shows that today's agents are a Vehicle 12 (a generative loop) wrapped in a Vehicle 5 harness (thresholds, tool calls, retries), with Vehicle 13 claimed on the box. And it keeps the attribution warning attached to the thing being sold.
Valentino Braitenberg, Vehicles: Experiments in Synthetic Psychology (MIT Press, 1984). Dates are first public release or publication; treat them as approximate where a product had a long beta. Companions: Vehicles, Drawn · Wiring the Week · How Agentic?