Appearance
Atlas
Give atlas a topic and an intent and it researches the topic wide-then-deep across the web, synthesizes what it finds into one structured bundle, and renders that bundle as a browsable, self-contained website. It is the research half of the harness productised into a single invocation. Where deep-research fans out searches and writes a flat cited report, atlas adds three things on top: intent-shaped modes, a structured findings bundle as the intermediate, and a generated site as the output instead of a wall of prose.
The engine is a budget-aware Workflow, so the fan-out is real parallel sub-agents, and the render is a self-contained html file that stands alone with no external assets. One is the expensive half and one is free and deterministic, and they meet at the bundle.
In Practice
/atlas <topic> research the topic and open the site
/atlas <topic> --intent find-libs shape the run, learn / find-similar / find-libs / map
/atlas <topic> --mode survey pick a preset, or let auto choose from the intent
/atlas <topic> --dry frame only, print the plan, spend no fan-out/atlas runs the Workflow, writes the returned bundle to a temp file, then renders it and opens the page. --dry is the cheap preview: it frames the run and prints the brief and the chosen preset without spending a single scout. A survey-scale run warns before it fans out, because that is the expensive one.
The Pipeline
Seven stages map onto the workflow phases, and the default path is a wide-then-deep funnel: scout the whole space, pinch to the best sources, then drill only those.
- Frame, one agent turns the topic and intent into a research brief, the sub-questions, one search angle per scout, and a short title for the run. It auto-selects a preset when the mode is unset, and runs a multi-perspective pass that surveys adjacent topics for angles the raw questions miss.
- Scout, B parallel searchers, one per angle, each returning scored candidate sources.
- Rank, a genuine barrier that merges every scout's candidates, dedupes by url and domain, and ranks by a credibility score before taking the top-K. This is the wide-to-narrow pinch.
- Deep, K parallel deep-divers, each doing a full fetch, extracting claims with citations, and following a few links. The compare presets score each candidate against a rubric here instead of writing prose.
- Verify, a gated adversarial pass that tries to refute each key claim and flags the unsupported ones. It is skipped on a compare run and under a budget floor.
- Synthesize, one high-effort agent merges the fragments into the bundle: ordered sections, a deduped bibliography, a comparison matrix, diagram specs, a glossary, related items, and the gaps the research could not close.
- Render, code, no agent. The bundle becomes the site.
Breadth and depth are the two dials, and named presets set them plus a framing. The budget clamps both down when it is tight, and every heavy stage watches budget.remaining().
The Modes
The presets are data, not code, so a new mode is a data edit. Each sets the breadth and depth dials plus a framing.
- funnel, scout wide then deep-dive the top-K, the sensible default.
- scout, a quick shallow survey. deep-dive, drill one known target. survey, the exhaustive map, the expensive one.
- compare, a matrix out, for find-similar and find-libs. ecosystem, diagram-led, map a space and its players. lineage, time-ordered history.
- auto, let the frame agent pick the preset from the topic and intent.
The Bundle
The structured intermediate every stage after synth reads. It carries the topic and a short title, a scannable summary, ordered sections each with a markdown body and cited claims, an optional comparison matrix, d2 diagram specs, a deduped source list that claims cite by index, a glossary, related items, and the gaps. Claims carry a verify verdict when the refute pass judged them, matched back onto the bundle after synth rewrites the prose. That carry is one for one, so a judged claim can never mark more than a single bundle claim, and where synth split one claim into several only the closest rewrite is marked. The contract is written out in full in the skill's bundle-schema.md.
The Render
The renderer turns one bundle into one self-contained html file, no external http of any kind, so it opens over file:/// and works offline. A left sidebar switches between content panes client-side, so a single file still reads as a multi-page docs site. Typed [LIB], [SEED], and [CAUTION] tokens the source controls become coloured badges, claims cite into the bibliography, verified and unverified verdicts show as styled spans, and d2 diagram specs render to inline SVG through the diagram toolkit, a wants dependency, with a code-block fallback when it is absent.
One look ships, fuma, a calm low-chrome docs style built from fumadocs own tokens, weight carrying the hierarchy so nothing reads thin on a system font stack. It styles light and dark, and uses system fonts only, because a webfont would break the offline guarantee. An earlier field-atlas look was retired on 2026-07-29 once fuma became the one Finn wanted.
Into a Wiki
The render stays one self-contained page, and a wiki can point at it rather than re-rendering the research as a native multi-page site. This is opt-in.
wiki-link.nu registers a generated page into a wiki: it records the page in the wiki's atlas-refs.nuon and stages it into docs/public/atlas/, so a running dev server shows it at once. The next scaffold gen, passed --references, renders a Research sidebar section and nav entry that link to each page, and scopes ignoreDeadLinks to /atlas/ so the build still fails loudly on any other broken link. Pass no references and the wiki is unchanged, which is what makes it opt-in.
Connects to
- Orchestration, the budget-aware Workflow patterns the engine is built on, the same shape as the orchestration scripts.
- Wiki, the docs site atlas links its generated pages into, through the reference-aware scaffold.
- Memory, where the lessons from building and running atlas are banked.