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

How work gets done

Five real requests, each walked through the system from the moment it arrives to the moment it changes what happens next. They take the same eight steps every time. Only what fills each step changes. Read two or three and the architecture stops being a diagram you have to trust and becomes something you watched happen. Two of the five work today, two are partly built, and the fifth does not work yet. The fifth is the most useful of the five.

This page is the one place in these docs that names files and tables. Every step below says which part of the system does it, so the walkthroughs and the architecture describe the same thing.

Work arrives through one of three doors: a message, a schedule, or an event. Whichever door it came through, it then follows one path. Two of the stations are easy to miss. The workflow is there because nobody ever picks a tool. The brand is there because it is what makes the same method produce different work.

#StationBuilt fromRecords
1Trigger
What starts the work: a message, a schedule, or an event?
web/src/app/api/chat/route.ts · tools/agent_chat_bridge/bridge.ts · tools/run_agent_cron.sh · tools/sync_crons.py · tools/ingest/cron_runner.pychat_threads · chat_messages · runs · signals
2Workflow
Which workflow does this activity run?
flows/ · tools/flow_progress.pyflows · decision_plays
3Context
What does the agent already know when it wakes?
tools/recall.py · tools/memory_write.pymemory_entries · wiki_docs · pipeline_tasks
4Skill
Which know-how loads, just before the step that needs it?
skills/ · tools/skill_runner.py · tools/wilson_mcp.pyskills · skill_invocations
5Brand
Whose voice, look and customers does the output follow?
brands/ · tools/brand_file.py · tools/elements.pybrands · elements · design_systems
6Action
Does it make something, pull a lever, or ask a person?
tools/media_generate.py · tools/amazon_ads_cycle.py · tools/agent_chat_bridge/bridge.tsgeneration_runs · action_levers · interventions · cockpit_settings
7Track
Where does the result live, and who decides it is good?
tools/pipeline_create.py · tools/append_task_version.py · tools/media_publish.py · tools/media_storage.py · web/src/components/workbench/goals · pipeline_tasks · task_versions · publish_operations · media_assets · task_reviews
8Learn
What changes next time, from the rating and from the outcome?
tools/recompute_skill_ratings.py · tools/showdown.py · tools/measure.pytask_reviews · skills · metric_snapshots · interventions · patterns
Three ways in, one path of eight stations. The rail is a labelled diagram; the table is generated from the architecture's concept map, and the build fails if a file or table in it stops existing.

The first and sixth stations vary the most, and they vary independently. A request can arrive through any of three doors and end in any of three kinds of action, which gives nine combinations on one path. The line from the last station back to the third is what makes this a system rather than a very good assistant.

Each of these is a real request, phrased the way it would actually be said. Pick one.

“Scout, give me three cover options for the holiday activity book.” works today

Asked in chat · Scout, the Code Pineapple agent · you decide at the end

  1. 01TriggerHow it knows to startYou type it into the chat panel on whatever screen you are on. The message starts a run in its own isolated workspace with a short-lived pass.web/src/app/api/chat/route.ts · tools/agent_chat_bridge/bridge.ts · chat_threads · chat_messages · runs
  2. 02WorkflowThere is always oneAn image task runs a workflow: brief, generate options, run the checks, present them. Scout names the workflow it chose.flows/ · tools/flow_progress.py · flows
  3. 03ContextWhat it already knewScout wakes knowing its open tasks, the decisions from last session and what you said about this book.tools/recall.py · memory_entries · pipeline_tasks
  4. 04SkillKnow-how, just in timeThe image skill loads immediately before the render step, not at the start of the conversation.skills/generate-images/ · tools/skill_runner.py · skills · skill_invocations
  5. 05BrandOn brandThe brand's palette, audience and customers' own phrases load at run time. That is the only reason this is a Code Pineapple cover.brands/code-pineapple/brand-essence.md · brands/code-pineapple/customer.md · tools/elements.py · brands · elements
  6. 06ActionMake, pull a lever, or askMake something. Three renders, each through the one gateway that holds the provider key and records a receipt before the render finishes.tools/media_generate.py · generation_runs · media_assets
  7. 07TrackVersion, task, goalThree versions on one task. The publish gate refuses to move the task to review until every file is registered.tools/append_task_version.py · tools/media_publish.py · pipeline_tasks · task_versions · publish_operations
  8. 08LearnRating and outcomeYou approve A and comment “warmer sky” on C. The rating is credited to the skill that made it, not to the agent in general.web/src/components/workbench/ · tools/recompute_skill_ratings.py · task_reviews · skills

Status checked against the running system on 2026-10-06. Each grey line names the files and tables behind that step; the build fails if one stops existing.

These are the principles, not because they were chosen as principles, but because they are what survived five walkthroughs that otherwise had nothing in common.

  1. There are only three ways in. A message, a schedule, an event. These are not three subsystems. They are three doors onto one recorded path, and the same rules apply whichever one a request uses. A fourth door would be a real architectural change.
  2. There is always a workflow. The activity selects the workflow, and the workflow says which skills run, in what order, with what checks. That is why a use case is the right unit to explain the system with: a use case is a workflow with a trigger on the front.
  3. Nothing starts cold. Every run began with context already loaded: the open tasks, the decisions from last time. A program assembles it when the agent wakes, kept lean on purpose, with the detail left until a step asks for it.
  4. Know-how arrives just in time, per step. The skill loads immediately before the step that needs it, not at the start. That keeps a session small enough to stay sharp while still being expert at the moment it matters.
  5. Brand is loaded, not baked in. The same skill makes a Code Pineapple cover or a RiseBlox one because the brand loads at run time. A skill never knows which brand it is for until it runs.
  6. An action is one of exactly three kinds. Make something, pull a lever in a system, or ask a person. This is the sharpest test of whether a request fits: if it needs a fourth kind of action, that is a genuine gap, and the email walkthrough is one.
  7. Money always stops and asks. Including inside a scheduled run in the middle of the night. The guardrail does not depend on anyone watching.
  8. Output is a version on a task, or it did not happen. Versions belong to tasks, and tasks serve goals. Anything that leaves the system unregistered is lost work.
  9. Every run ends by changing what the next one knows. A rating lands on the version it was given on and scores the skill that made it. An outcome is compared with its forecast. Over time, a skill or play that has earned it should move up the trust ladder, from propose only to just handle it.

The walkthroughs are also an audit. These are the places where the path is designed but not yet carrying weight. What works today holds the current status of each.

  • Track has a hole in the middle. Versions, tasks and goals are real records. Projects are not: a task carries a project name as a label, and there is no project record to group them. A task can point at a goal, but nothing works out which goal a task moves.
  • Learn has one input of two. Ratings score skills today. The outcome input, where a forecast is compared with what happened and the result shapes the next forecast, has its fields and no writer. Moving a skill or a play up the trust ladder is a rule followed by hand, not something the system computes.
  • Acting alone is switched off. The advertising cycle proposes changes; it does not pull levers by itself yet.
  • The event door is shaped like a social post. An event record requires a post, a link and a view count, so anything that is not a post (an email, a booking, a payment) has nowhere to land. The email walkthrough shows what that costs.
  • There is no fourth action. Reaching you outside the app is limited to a scheduled email and the coding factory’s phone notifications. An agent cannot choose to do it.

Sources: How Work Gets Done (walkthrough page, 2026-10-02), absorbed and re-checked against the running system 2026-10-06, docs/DECISIONS.md (ADR-029, ADR-028), docs/architecture/overview.md (the loop), docs/architecture/concept-map.json (stations), docs/architecture/engine.md · Last verified 2026-10-06