We build agents, visual n8n workflows and review interfaces around how you work. We develop the guidance and skills with you, then test real examples together so you can see what ran and judge the result.
See the exampleThe Living Article brings together research, different perspectives, article writing and reader discussion. It gives us a real system through which to explain our work.
An n8n Agent uses skills and calls documented workflows. Skills hold reusable instructions and tools; scripts handle suitable calculations. The visual structure helps you find a step, then inspect what actually happened.
Pattern Works’ own development project, using Supabase storage and a custom reader. It is a different article from the historical investigations below. For other work, we can build Airtable interfaces that let people review and correct workflow outputs.
Explore an actual output
Open “Understandable AI Is Not One Interface” to read the saved article, follow its sources and expand the research behind four AI personas. This is Pattern Works’ own development project.
Open a persona’s research, follow a source and compare what it supports with the article’s argument. The personas are AI-generated viewpoints, rather than human researchers or client testimonials. The historical investigations below concern separate runs and show how we traced particular engineering problems.
We work through real cases with you, capture exceptions and decide what belongs in rules, retrievable knowledge and skills. We help make that guidance usable; you bring the experience needed to judge it and identify decisions that should stay with a person.
We build the agents, skills, connected workflows and review interface needed for the task. Airtable can put the outputs in a queue where you check, correct and approve them. We use your feedback and representative examples to test changes.
The orchestrating n8n Agent has skills for its tasks and workflow tools it can call. Each workflow has a purpose, inputs and a result you can inspect. The records help you question an answer or investigate a failure.
A person can use an investigating agent, guided by skills, to help find the relevant execution. Its explanation should lead back to records the person can check.
One earlier article completed successfully but produced only one short body section. Inspection showed that the outline had left the planned material beneath a shared parent heading, so the writer scheduled one section.
A corrected outline test produced eight independent writing tasks. A later fresh article had eight body sections. Those are separate recorded runs: useful evidence about the structure, with article quality still requiring review.
What the investigation connected
The short article. Its saved outline. The section plan. The correction and its checks.
Historical example from 16–17 October 2026. No live investigation or conversation is being simulated on this page.
The Living Article gives us actual work to inspect. Its recorded investigations can help identify checks worth adding to future runs.
The outline left planned material beneath a shared parent heading, so the writer scheduled one section. A corrected outline test produced eight writing tasks. A later article had eight body sections.
Separate recorded runs, 16–17 October 2026.
Compare the intended section plan with the writing tasks before generation. A missing task can be flagged before the writer starts.
Proposed check. Passing it would not establish the quality of the article.
Assess the argument, evidence and other agreed criteria across the complete output. Review the mark and recommendations, and refine the metric where the assessment misses something.
A general article metric and judge remain in development.
Living Article’s feedback reviewer uses Jev to assess reader concerns. Recorded tests show suggestions being grouped and saved.
That provides a useful pattern for collecting findings and preparing proposals for review. It does not yet demonstrate a general system that automatically improves any workflow.
How we approach evaluationsStart with one useful piece of work, or bring a clear implementation brief. We’ll agree what to deliver, the examples and expert time needed to assess it, and what happens after handover.