Skip to content

Roadmap v2 — High-Level Backlog

Field Value
Version v2.0
Status PLANNED (not week-scheduled)
Date March 2026
Parent PRD-001 Master Overview

This document captures the themes and features explicitly deferred from v1.0 plus natural extension areas identified during v1 design. v2 items are not week-scheduled; they will be prioritized after v1.0 ships and user feedback is collected.

Theme 1 — Multi-Tracker Support

v1 scope: GitHub issues only.

v2 extension: Support additional issue trackers as first-class input sources.

  • Jira Cloud: OAuth 2.0 integration, issue fetch via Jira REST API v3, write-back via POST /issue/{key}/comment
  • Linear: GraphQL API integration, status sync, team/project scoping
  • GitLab: Issues API, MR linking, self-hosted GitLab instance support
  • Generic webhook ingest: any tracker that can POST a structured payload gets routed through the agent graph

Key design decisions needed:

  • Abstraction layer (IssueSource interface) so agents receive a normalized IssueContext regardless of origin
  • Auth credential management per workspace (each user may have Jira + GitHub credentials simultaneously)
  • Write-back adapter per tracker (comment format differs)

Theme 2 — Non-Software Domains

v1 scope: Software bug triage only.

v2 extension: Generalize the agent graph to support other structured investigation workflows.

  • Legal document review: Investigator reads contract clause, Critic identifies risk, Writer drafts summary
  • Security incident triage: Alert ingested from SIEM, agents research CVEs, recommend remediation
  • Customer support escalation: Ticket from Zendesk/Intercom, agents search knowledge base, draft reply
  • Data quality investigation: Anomaly detected in pipeline, agents trace data lineage, identify root cause

Key design decisions needed:

  • Domain-specific agent prompts and output schemas per domain (plugin system — see Theme 7)
  • Domain-specific tool sets (legal DB search, CVE DB, knowledge base RAG)
  • UI output panel templates per domain

Theme 3 — Self-Hosted LLM Support

v1 scope: OpenAI and Anthropic cloud APIs only.

v2 extension: Allow users to run agents against locally-hosted or self-hosted models.

  • Ollama: Local model runner; expose as OpenAI-compatible endpoint for drop-in use
  • vLLM: High-throughput inference server for enterprise self-hosted deployments
  • LM Studio: Desktop app for Mac/Windows; compatible via local OpenAI API
  • Custom fine-tuned models: Support base_url override per agent in config

Key design decisions needed:

  • Per-agent model config (each agent may use a different model or backend)
  • Latency and context-window constraints for smaller local models (adjust prompts accordingly)
  • Fallback chain: if local model times out, optionally fall back to cloud API

Theme 4 — Real-Time Multi-User Sessions

v1 scope: Single-user session; one job owner.

v2 extension: Multiple users can observe and interact with the same job simultaneously.

  • Shared job view: multiple users see the same live workspace in real time
  • Collaborative question answering: any permitted user can answer an agent's question
  • Role-based permissions: owner (full control), reviewer (can answer questions, cannot kill), observer (read-only)
  • Presence indicators: show which users are currently viewing a job (WebSocket-based)
  • Conflict resolution: if two users submit an answer simultaneously, use first-write-wins with notification to the other

Key design decisions needed:

  • WebSocket upgrade for presence and collaborative state (vs. SSE which is unidirectional)
  • Session isolation: each job has an ACL; sharing requires explicit invite
  • Audit log: every user action on a shared job is recorded

Theme 5 — Mobile Interface

v1 scope: Desktop web app only (1280px+ viewport).

v2 extension: Native-quality mobile experience for monitoring and responding to agent questions on the go.

  • Responsive layout: job queue as bottom sheet, workspace as full-screen view
  • Push notifications: notify job owner when an agent asks a question or a job completes
  • Answer-on-mobile: question card is thumb-friendly with large input area and confirmation step
  • Mobile-first output panel: collapsible sections for each output field
  • PWA packaging: installable from browser on iOS and Android

Key design decisions needed:

  • Native app vs. PWA (PWA preferred for v2 to avoid App Store release process)
  • Push notification delivery: Firebase Cloud Messaging or Apple Push Notifications
  • Reduced streaming fidelity on mobile (batch chunks to reduce re-renders on slower connections)

Theme 6 — Expanded Evaluation & A/B Prompt Framework

v1 scope: Single LangSmith eval dataset, manual evaluator script.

v2 extension: First-class evaluation and experimentation infrastructure for continuous agent improvement.

  • A/B prompt testing: Define prompt variants per agent; route a % of jobs to each variant; compare LangSmith eval scores
  • Automated regression suite: CI runs evals on every PR against the bug_triage_v1 dataset; blocks merge if score drops > 5%
  • Human feedback loop: Thumbs up/down on final output in UI; feedback stored as LangSmith annotation; auto-added to eval dataset
  • Cost/quality Pareto dashboard: Plot each agent configuration on a cost vs. quality 2D chart; surface Pareto-optimal configs
  • Shadow mode: Run new agent versions on live traffic without affecting the user-facing response; compare outputs asynchronously

Key design decisions needed:

  • Prompt version registry (link prompt hash to eval result in LangSmith)
  • Sampling strategy for A/B (per-user consistent assignment vs. per-job random)
  • Feedback data governance (PII scrubbing before adding to eval dataset)

Theme 7 — Plugin / Extension System for Custom Agents

v1 scope: 5 fixed agents (Investigator, Codebase Searcher, Web Searcher, Critic, Writer).

v2 extension: Allow teams to add custom agents and tools without modifying core code.

  • Agent plugin interface: Define an AgentPlugin protocol (Python); any class implementing it registers as a new agent node
  • Tool plugin interface: Define a ToolPlugin protocol; custom tools are injected into existing agents via config
  • Plugin registry: UI panel to browse, enable/disable, and configure installed plugins
  • LangFlow import: A LangFlow flow file can be registered as a plugin agent; no Python code required
  • Plugin marketplace (future): Public registry of community agent plugins (similar to VS Code extensions)
  • Sandbox execution: Plugins run in isolated subprocess or container to prevent host compromise

Key design decisions needed:

  • Plugin discovery mechanism (file system scan vs. entry points vs. explicit config)
  • Versioning and compatibility: plugin declares compatible core version range
  • Security model for sandboxed plugins (what resources can a plugin access?)
  • Plugin state access: read-only vs. read-write AgentState

Deferred Non-Goals from v1 — Resolution in v2

The following items from the v1 Non-Goals are formally tracked here:

v1 Non-Goal v2 Theme
Non-GitHub tracker support Theme 1
Non-software domains Theme 2
Self-hosted LLM support (Ollama etc.) Theme 3
Real-time multi-user collaborative sessions Theme 4
Mobile interface Theme 5
Fully autonomous operation (zero oversight) Theme 6 (eval-gated)