Roadmap#

Where the platform stands and where it goes next. Sequenced by dependency and payoff, not wishful thinking — items reference the docs that explain the gap.

Where we are (shipped)#

• ✅ Agent CRUD (create/read), single-turn chat with run telemetry
• ✅ Email analysis with task extraction; task status management
• ✅ Four-stage multi-agent email workflow with full tracing and quality scoring
• ✅ RAG: .txt ingestion → Chroma → corrective LangGraph answering, per-agent isolation
• ✅ Observability: ToolCall/AgentRun/WorkflowRun/WorkflowStep + dashboard aggregation
• ✅ Next.js dashboard, agents list, create-agent UI with proxy pattern
• ✅ This documentation, deployed on Vercel

Milestone 1 — Correctness & completeness (next)#

The gaps that bite users today; most map to Known Issues:

• Fix/remove the broken stats endpoint; ship real PUT /agents/{id} and DELETE /agents/{id} (with Chroma vector cleanup)
• Fix the generate_rag_answer loop-indentation landmine
• Close the ingestion partial-failure window; add upload size limits
• Uniform cost logging across chat / analyze / RAG / embeddings (coverage map)
• Real token counts from response.usage; numeric cost/score columns
• Conversation memory — replay recent Message history into chat context (the gap)
• Test foundation: Tier 1–2 from the testing plan

Milestone 2 — Frontend depth#

• Agent detail page (/agents/{id}): chat UI, runs, tool calls, tasks, documents
• Dashboard: agent picker (kill the agent-1 pin); real chart series from a time-bucketed endpoint
• Wire the sidebar's Workflows/Operations sections to real pages
• Proxy routes for chat/documents/workflows following the pattern
• Dark mode for the app (the docs have one; the app doesn't)

Milestone 3 — Production hardening#

Prerequisite for any non-local deployment (Security):

• Authentication (API keys/JWT) + agent ownership; CORS middleware
• Rate limiting and a per-agent OpenAI spend circuit breaker (the cost data exists to build on)
• Retries with backoff at the OpenAI boundary; sanitized error details
• Dockerfiles + compose; CI (typecheck, tests, mkdocs build --strict)
• Structured logging; /health uptime monitoring

Milestone 4 — Scale & platform#

• Postgres via Alembic (graduation criteria); pgvector or hosted Chroma
• Background execution for workflows (queue + status polling) — unlocks long pipelines and retries (current constraint)
• Streaming chat responses; batch embeddings on upload
• More ingestion formats (PDF/DOCX); document deletion & re-indexing
• Full CRAG retry loop (query rewrite on refusal — sketch)
• Second workflow type; workflow definitions beyond hardcoded pipelines
• Eval harness over recorded traces (Tier 5)

Non-goals (for now)#

Multi-region deployment, fine-tuning, non-OpenAI model providers, real-time collaboration, and a plugin marketplace — revisit after Milestone 3.

How to influence this#

Open an issue or PR against this page — the roadmap is a living document, and entries should always link to the docs explaining the current state (Contributing).