Skip to content
Describes how WilsonOS is designed to work.What works today →

Engines

Pillar 05 · Engines

The repeating parts of the business. Each cycle an engine predicts what will work, does it, and checks what happened — that gap is what teaches it.

A channel that runs on a cycle, with you approving the plan. Advertising, Pinterest, YouTube, short video — each was once going to be its own dashboard with its own logic. Read side by side, they turned out to be the same machine. Every engine runs one cycle, and because its feedback is clear and fast, engines are where agents can most safely be trusted to run more on their own.

One cycle, every engine: aim at a number, write down a forecast, act, compare, and let the gap sharpen the next forecast.
What you do
  • Set the number an engine is moving.
  • Decide how much it may do on its own, and its spend caps.
  • Approve what crosses your limits, from its cockpit or a conversation.
What your agent does
  • Writes down what it expects before it acts.
  • Records what it intended and what was actually applied.
  • Scores its forecast against what happened, and says how often it is right.
  • Never approves its own work.
  1. Target — the goal it is moving, as a number.
  2. Predict — before acting, it writes down what it expects a change to do.
  3. Act — it makes something or pulls a lever, and records what it intended and what was actually applied.
  4. Measure — it compares what happened with what it predicted.
  5. Learn — the gap sharpens the next prediction. That is why it can say “2.1× your normal save rate, and it’s the title treatment, not the subject” instead of “this did well”.
  6. Rope — how much it may do on its own: limits, a kill switch and spend caps. It runs its own steps and stops at yours.

There are three types of engine on that one contract: advertising, which tunes bids and budgets on what is already live; channel, which makes new creative and tunes what is live; and delivery, which predicts a project’s scope against hours (designed, not built).

Reach Engage Convert
Advertising Impressions Clicks Orders
Pinterest Impressions Saves Outbound clicks
YouTube Impressions Views Watch-through
Short video Reach Watch-through Saves / follows

Only the metric names and the levers change. So the engine is built once, and each channel plugs in a domain pack: its metrics, its levers, what good creative looks like there (a pin, a thumbnail, a hook), and its playbook. A create-and-tune channel makes new creative every cycle as well as adjusting what is live; a tune-only channel such as advertising has nothing to make and adjusts existing keywords, bids and budgets.

Every step produces the same thing: a decision, a forecast plus an action. Decisions land in one ledger, and next cycle each forecast is scored against what actually happened. That calibration — how often the engine’s predictions come true — is the spine of “getting smarter” on a channel. Ideas come from several places, and not all of them are data: a trend promoted from research, a batch for a destination (“twelve pins for this lead magnet”), a gap no one has tried yet, or you saying “make this”.

Each engine has a cockpit, and they share their shape: an overview, the signals it is watching, ideas, the plan, the patterns that are proven, the data, and the playbook. Learn one channel and you know them all; a tune-only channel simply greys out the making stages. The intent is that turning on a new engine gives you its cockpit from the same template. Engines use the shared Workbench for review and never get a review screen of their own.

An engine mostly runs on a schedule: it pulls the channel’s signals, works out what changed, and proposes. Proposals arrive in a conversation like anything else. Making creative goes through the usual workflows and review; adjusting something live is an action through the channel’s connection, with its permissions. Nothing posts, pauses or spends without the decision being visible.

Question Scope Answered by
Is this channel converting? Inside one channel: reach → engage → convert That channel’s engine
Where is the bottleneck to more sales? Across channels: social → landing page → lead → offer → sale Funnel tracking
Are we making money? Royalties, advertising, other costs Business performance, in the shared brain

They chain — channel feeds funnel feeds business — but they answer different questions at different paces. Treating “analytics” as one pile is how a dashboard ends up confident and wrong.

Attribution limits are stated, not hidden. The system shows what it can prove (“forty-one visits from the sales page, six purchases”) and says what it cannot (“a pin impression cannot be linked to this store purchase”). A brand’s funnel steps currently live in four systems that share no identifier; the piece that joins them is a tracked destination link the brand owns, and until a path has one, that path is reported as not yet measurable rather than guessed.

Results go two places: to you, for the next decision, with source and freshness beside every figure; and to the improvement pass, alongside your ratings but in a separate column, because “the owner liked it” and “it worked” are different facts.

For builders — the decision ledger and the loop
  • One ledger of interventions. Every decision is stored as a forecast plus the action intended and the action actually applied, so a cycle can be scored even when what was applied differed from the plan.
  • Scheduled pulls, scored forecasts. Signals arrive through the shared ingestion pattern on a registered schedule; an engine without a schedule has a snapshot, not a feed, and cannot calibrate.
  • The agent’s loop, smaller and faster. An engine runs the same shape as the agent loop inside one pillar. Engines supply the outcome half of learning; your ratings supply the other half.
  • Rope is configuration, not trust by default. Limits, the kill switch and spend caps are set per engine. Moving an engine up the trust ladder — from proposing only, to acting and reporting, to just handling it — is designed, not built.

Sources: docs/architecture/engine.md (six steps, three types), docs/architecture/channel-engine.md (three gates, two archetypes, domain packs, decision ledger), docs/architecture/measure.md (three questions, attribution limits, tracked destination links), docs/DECISIONS.md (ADR-028), Don's product direction, 2026-09-16 (task 410 review) · Last verified 2026-10-06