General Bots Roadmap 2018-2027 🟑 BETA


Last verified September 2026. This page states what has shipped, what is in progress and what is planned. Every β€œshipped” entry is traceable to code in this repository. The interactive view above is the 2018–2026 planning artifact and is not maintained as a status source β€” this page is.

Current release state

SurfaceStatus
Chat, Explorer (Drive), VibeStable
Mail, SheetsPreview β€” advanced
Docs, SlidesPreview β€” in test
Everything else in the catalogPreview

Per-application state, with counts generated from the catalog, is in Suite Apps Status.

Timeline

β€œDelivered” counts items actually present in the code today, not items that were scoped.

PeriodFocusDeliveredState
2018–2024v1–v5, pre-LLM12βœ… Shipped
2024v6 foundation β€” Rust core, PostgreSQL, Vault8βœ… Shipped
2025 H1Rust migration β€” BASIC engine, channels, Drive (S3), email, REST, WhatsApp, Telegram, PDF10βœ… Shipped
2025 H2Features and Autotask β€” tasks AI, knowledge base, vector search, tools, generators, multimodal, NVIDIA GPU/LXC, Paper, Research, Calendar, Meet19βœ… Shipped
2026 Q1Autonomous tasks, cloud productivity connectionssee below🟑 Partially shipped
2026 Q2Collaboration and multi-agentsee below🟑 Partially shipped
2026 Q3Workflow and CRM β€” Designer, CRM2βœ… Shipped
2026 Q4Enterprise β€” mobile apps, enterprise SSO, white label, advanced monitoringβ€”πŸ”΅ In progress
2027 H1Retrieval quality β€” evaluation harness, re-ranking, chunkingβ€”πŸ“‹ Planned
2027 H2Retrieval depth β€” graph index, multi-hop verificationβ€”πŸ“‹ Planned

2026 detail β€” what is actually there

ItemStateEvidence
Multi-agent orchestrationβœ… Shippedbotserver/src/core/bot/ agent pipeline, docs
Cloud productivity connections (Google, Outlook, OneDrive, Google Calendar)βœ… Shipped as integration providersbotserver/crates/botintegrations/src/providers/
Messaging providers (Slack, Discord, Teams, Zoom)βœ… Shipped as integration providerssame provider catalog
White labelβœ… Shippedthe .product configuration surface, docs
Enterprise SSOβœ… ShippedZitadel-backed identity, botcoredirectory
Advanced monitoringβœ… ShippedMonitoring app, botmonitoring
App marketplaceβœ… Shippedbotmarketplace, App Store app
Workflow designerβœ… ShippedDesigner app
CRMβœ… ShippedCRM, People, Sales and Campaigns apps
Mobile appsπŸ”΅ In progressTauri shell exists for desktop; mobile not verified
Fully autonomous task execution⚠️ PartiallyAutotask runs, but β€œproduction autonomous” as originally scoped is not verifiable as complete β€” treat as in progress

2027 horizon β€” planned

Ordered by what unlocks the most, not by what is most visible:

  1. Retrieval evaluation harness. Without a golden set and a measured metric, every other retrieval decision is guesswork. This blocks the rest β€” see #1379.
  2. Cross-encoder re-ranking. The cheapest quality win available; the configuration surface already exists in an unconnected crate.
  3. Real chunking with overlap. Documents are currently indexed whole, which costs precision on long documents.
  4. Graph retrieval with an actual graph index, replacing entity expansion with traversal.
  5. Latency and cost budgets per retrieval mode, so mode selection stops being a guess.

Technology stack β€” verified

LayerTechnology
LanguageRust
HTTPAxum (with axum-server)
Async runtimeTokio
DatabasePostgreSQL with Diesel and diesel_migrations
CacheValkey
Object storageMinIO
Vector searchQdrant
IdentityZitadel
UIHTMX with Askama templates
DesktopTauri

Correction. Earlier revisions of this page listed Actix-Web and SQLx. Both were wrong: the repository contains no Actix dependency and no SQLx usage β€” the HTTP stack is Axum and the data layer is Diesel.

Status legend

StatusMeaning
βœ… ShippedPresent in the code and reachable by a user today
🟑 Partially shippedSome of the scope works; the rest does not
πŸ”΅ In progressActive work, not yet reachable end to end
πŸ“‹ PlannedScoped, not started
⚠️ Not verifiableClaimed previously without evidence β€” do not rely on it

How this page is maintained

  • A row only says β€œshipped” if someone traced it to code.
  • Counts that can be derived from the repository are generated by scripts in scripts/, not typed by hand.
  • When a plan changes, the old state is corrected rather than deleted, so the record stays honest.