Roadmap
The build queue, in the open.
Ten milestones, M0 to M9. Many agents work through the queue in parallel, so it moves in days of work. Counts are features done out of features planned.
Updated
Milestones
- M0
Ground truth
Measure everything. Nothing counts as done without a receipt. Task claims and rulings are written down.
23 / 30 done
- M1a
Security and trust floor
Rate limits, handlers that fail closed, and screening for prompt injection.
1 / 13 done
- M1b
Accounts and upgrades
One-click Pro, account delete and export in every store, consent, an age check and honest copy.
1 / 35 done
- M1c
Test debt to zero
Every known failing or missing test fixed.
5 / 7 done
- M2ui
Bubble Glass
Grouped bubbles, delivery status, jump to latest, paste and drop, What's new and in-app help.
0 / 24 done
- M2a
Fusion Gateway and Gems
One endpoint, routing per request, and Gems you can pick.
0 / 26 done
- M2b
Products
Tools with approval cards (Calendar first, then Gmail), voice, developer docs and an SDK, and memory controls.
0 / 49 done
- M2c
Messaging
Telegram and Slack first, approval by text, and daily caps.
0 / 23 done
- M3
AI Factory and enterprise
Seven departments on real providers, department work orders, and enterprise deploys on Cloudflare.
1 / 33 done
- M4
Research paper
An honesty revision, preregistration and benchmark integrity.
1 / 15 done
- M5
Paper to product
Design docs, the whole map, and evidence behind every model claim.
0 / 18 done
- M6
Self-improvement loops
Conversations and models that defend, heal and improve themselves. Each loop is preregistered and gated.
0 / 21 done
- M7
Brand truth
A claims audit. juwel.ai is for JUWEL and vextlabs.ai is for the lab.
0 / 23 done
- M8
Distribution
SDK and CLI packs, a VS Code extension, and mods.
0 / 20 done
- M9
Agent runtime
Durable task streams, budgets and hard stops, undo, a flight recorder, a support desk, a marketplace and the rest of a 44-item list.
0 / 55 done