Anaconda Next is where we test the edges before they become products — playable games about the package ecosystem, interactive field guides that make hard ideas tangible, and agent research with measurable evidence — published the moment it's ready to be public, never before.
An interactive Python field guide to map projections: Maps with Python. Grab a globe, flatten it, and watch every projection bend the truth differently — all computed in your browser.
A sailor needs a steady bearing. A reader comparing countries needs fair areas. Every flat map is an optimization problem, so we built the tradeoffs you can feel: drag a real globe, then watch Tissot's distortion circles stretch across Mercator, Gall–Peters, Equal Earth, and AuthaGraph.
Everything runs locally in your browser — the geodesy is computed by NumPy in Pyodide, the globe is rendered with Three.js, and every projection's math is cited down to the paper.
A daily game about the packages that make up the Conda and Python ecosystems. One clue at a time, from obscure package history down to a final mercy clue — wrapped in a lovingly recreated late-1980s DOS interface.
A daily clue-ladder puzzle about real conda packages, fact-checked against Distro-DB records, project docs and maintainer lore. Five clues per package; the modal solve lands at clue 2–3. How deep can you go?
Can an agent that investigates data through a live, replayable nteract notebook — execute a small step, observe the real result, retain state — solve analytical tasks more reliably than an answer-only agent? We're measuring it.
Give an agent a quantitative eval up front and let it optimize its own harness — tool schemas, retry policy, turn budget — not just the prompt. Inspired by WHALE (2026); the question is whether the gain survives held-out tasks.
Local agent quality is a property of model × quantization × llama.cpp settings × harness × hardware. We benchmark all five together on real tasks so Desktop can ship configurations that actually work.
Can semantic search help an agent find the right conda package from a task description? Across 42 synthetic queries and 37,801 packages: 62% top-10 baseline, 83% with popularity reranking — and a bias we don't want to ship.
A focused Neovim integration for Kilo's editor protocol — terminal toggling, placement, file mentions, selected code — hardened against real terminal edge cases and handed to the Kilo organization as an official plugin.
An agentic onboarding MVP spec with explicit safety gates, a two-week falsification sprint, and runbooks whose executor cannot self-certify success. Lab results may never be dressed up as employee-ready.
The reasoning behind the artifacts — what we tried, what we measured, and what we'd do differently. New pieces land here as the work earns a public story.
Baseline retrieval found the expected conda package in the top 10 for 62% of task queries. LLM enrichment moved the number exactly nowhere. The full evaluation story — and why 83% with popularity reranking was a number we refused to ship — including the biases we caught in our own expected-answer lists.
Read the write-up ▸ September 2026GitHub's Advisory Database teaches an agent to avoid packages already known to be dangerous. It cannot tell an agent that an unknown package is trustworthy. Notes on where controlled package admission fits beside vulnerability intelligence, and the three integration paths that emerged.
Read the write-up ▸ September 2026Real geodesy, no install: NumPy running in Pyodide, a Three.js globe, Natural Earth geometry, and distortion circles after Tissot's 1858 idea. How the projection math, the data pipeline, and the "no Python setup needed" promise came together — sources cited all the way down.
Read the write-up ▸ September 2026Five clues per package, ordered as a single information gradient, with the modal solve landing at clue two or three. How we fact-check every rung against Distro-DB records, project history, and maintainer lore — and decide which famous stories not to use.
Read the write-up ▸Why we think an agent that executes a small step, observes the real result, and leaves behind a replayable notebook beats one that just answers — and how we're testing that with deterministic scoring.
Drafting…The operating formula behind every run of this team: why prototypes stay subordinate to decisions, why negative results are published first-class, and what changes when you separate evidence generation from production ownership.
Drafting…Every incubation run follows one operating formula. Prototypes stay subordinate to decisions, negative results are first-class, and evidence generation stays separate from production ownership.
Proved useful in real terminals, hardened for edge cases, adopted upstream. The Pi rewrite question was killed fast: feasible, but no meaningful risk reduction over the OpenCode path.
Quizaconda, Condary, and Maps with Python went from inside jokes to cited, single-file, zero-dependency artifacts anyone can open. Public evidence doesn't have to wait for product readiness.
LLM enrichment added nothing; popularity reranking amplified bias. We shipped known-package lookup and deferred discovery rather than shipping weak retrieval. The evidence for revisiting is spelled out.
We test our thesis — that agents deserve affirmatively trusted environments, not just vulnerability warnings — in conversations with the people building the platforms agents will live on.
GitHub's Advisory Database helps an agent avoid known bad packages. Anaconda's controlled intake, independent builds, and malware scanning establish that an unknown package is trusted. Two different layers of the same defense.
Three paths surfaced from the conversation: an Anaconda-maintained base image for GitHub's emerging cloud-agent
microVM runtime, a setup-anaconda Action for approved channels on Actions runners, and Anaconda MCP + skills in GitHub's plugin marketplace.
Metaflow orchestration has run against Nebius compute pools since March 2025, with capacity-block pools and perimeter-scoped reserved pools now self-serve. The evidence pack we compiled ranks what customers actually need: a second GPU source without re-plumbing, and honest operational transparency.
The same evidence names the gaps — environment reproducibility, cost attribution, supply-chain governance for whatever runs on the GPU. That's exactly where the incubation bets aim.
“This format keeps prototypes subordinate to decisions, makes negative results visible, and separates evidence generation from production ownership.”