Autonomous Multi-Agent Triage Pipeline with Hermes Kanban

Hermes Agent v2.0+ runs a multi-agent triage pipeline that detects signals from multiple sources (Discord, GitHub issues, web mentions), deduplicates them against a shared SQLite kanban board, scores each item against a configurable rubric, researches high-scoring items in parallel using Tavily web search and GitHub MCP servers, routes each to a fulfillment path (bug fix, feature request, content piece), pauses at a single human approval gate, then fulfills and delivers the output through Telegram or Discord gateways. The agentic reasoning step involves the orchestrator agent comparing parallel research findings from subagents, detecting contradictions between sources, and requesting targeted re-research before assigning a composite confidence score. Measurable outcome: 100+ signals triaged daily with under 5 minutes of human review time per accepted item.

BUSINESS PROBLEM

A 5-person growth team at a B2B SaaS company monitors 8 channels for user feedback: Discord community, GitHub issues, Twitter mentions, support tickets, NPS responses, app store reviews, Reddit threads, and Intercom chats. Each team member spends 2-3 hours daily manually scanning these sources, identifying duplicates, and deciding what to act on. [ STAT ] 68% of product teams report that signal triage consumes more time than actual feature development, with the average team missing 40% of actionable signals due to channel fragmentation. — ProductPlan Signal Management Report, 2025. The team cannot hire more people, so signals that fall through the cracks mean lost revenue opportunities, unresolved user complaints that escalate to churn, and competitive intelligence blind spots that lead to incorrect product decisions. The manual triage process introduces a 24-48 hour latency between signal arrival and action, during which users may abandon the product or competitors respond first.

WHO BENEFITS

1. Growth teams at B2B SaaS companies who monitor 5+ user feedback channels and need an automated triage system that surfaces high-signal items (feature requests with 10+ upvotes, bug reports affecting paying customers) within minutes rather than days of the signal appearing in any monitored channel. 2. Product operations managers at mid-market companies who run weekly backlog grooming sessions and want a pre-scored, researched, and routed candidate list ready for review every Monday morning, cutting grooming time from 4 hours to 45 minutes per week across the entire product team. 3. Developer relations engineers at API-first platforms who manage a community of 10,000+ developers across Discord, GitHub Discussions, and Stack Overflow and need to automatically detect unanswered questions, API usability friction signals, and documentation gaps buried in hundreds of daily messages.

HOW IT WORKS

1. [TOOL: Hermes Agent Gateway] Signal intake: Hermes runs a cron job every 15 minutes that executes scout agents across configured sources. Input: 8 channel configurations (Discord channels, GitHub issue queries, Reddit subreddit streams). Output: raw signal items with source metadata, timestamp, and urgency heuristic score. 2. [TOOL: Hermes Kanban Board] Deduplication: the intake agent checks each new signal against the SQLite kanban board's existing items using token-cosine similarity. Duplicates below a 0.85 threshold are merged with the existing item's signal count incremented. Input: raw signal item. Output: new kanban card or updated existing card with incremented signal strength. 3. [TOOL: Hermes Agent Orchestrator] Rubric scoring: the orchestrator agent scores each item against a configurable rubric defined in triage.yaml with criteria including user tier (paid vs free), signal frequency (first report vs 10th), revenue impact estimate, and alignment with current sprint goals. This is the AI reasoning/decision point: the agent weights each criterion dynamically based on the current business context read from the kanban board metadata, so a bug affecting a recently shipped feature scores higher than the same bug in a legacy module. 4. [TOOL: Tavily MCP Server] Parallel research: items scoring above 70 are dispatched to parallel research subagents. Each subagent runs a Tavily web search for the topic, fetches top 5 sources, and extracts key findings into a structured research brief. Input: scored kanban card with item description. Output: research brief with source URLs, key findings, and suggested action. 5. [TOOL: Hermes Subagent System] Fulfillment routing: the orchestrator classifies each researched item into a path: bug fix routes to GitHub issue creation, feature request routes to a spec draft, content piece routes to a blog outline. The routing decision is logged in the kanban card. 6. Human Approval: the orchestrator posts a daily digest to a Telegram group with one card per item showing score, research brief, suggested route, and an Approve/Shelve/Modify button. The human reviews all items in under 5 minutes. This is the single human approval gate. 7. [TOOL: GitHub MCP Server] Fulfillment: for approved items, the pipeline creates a GitHub issue with the research brief as the description, assigns a sprint label based on urgency, and posts the issue link back to the kanban card. Output: created GitHub issue with structured body. 8. [TOOL: Hermes Cron] Delivery: the fulfillment agent posts a summary to Discord or Telegram showing what was triaged, what was approved, and links to created artifacts. Input: fulfillment results. Output: platform message with item links and next-steps call to action.

TOOL INTEGRATION

Hermes Agent v2.0+: Install via pip install hermes-agent or the Nix flake. Configure profiles in ~/.hermes/profiles/triage/ with isolated memory and skills. The orchestrator runs as the default profile; research subagents use a research-profile with Tavily MCP enabled. Gotcha: Hermes Agent profiles share the same filesystem but have separate SQLite state databases. If a subagent crashes mid-research, its partially written findings are lost unless you configure a checkpoint directory in the profile config at ~/.hermes/profiles/triage/config.yaml under checkpoint_dir. Hermes Kanban Board: Built on SQLite with a file-based lock. Initialize with hermes kanban init --board triage. The board stores items in kanban_tasks.db with columns for id, status, dependencies, and results. Gotcha: SQLite's default WAL mode can cause file lock contention when 5+ subagents write simultaneously. Set pragma journal_mode=MEMORY and use a dedicated write queue with retry logic in the orchestrator hook. Tavily MCP Server: Install via npx @tavily/mcp-server and configure in hermes config set mcp_servers.tavily. The server provides web_search and web_extract tools. Rate limit is 1000 requests/month on the free tier. Gotcha: Tavily's extract function has a 10-second timeout for slow pages. Configure a retry with exponential backoff in the research subagent's SOUL.md using retry_on_timeout: 3. GitHub MCP Server: Install via npx @anthropic/mcp-github. Scope the token to contents:read, issues:write, and pull_requests:write. Gotcha: GitHub MCP server uses the GraphQL API for issue creation, which has a different rate limit than REST API. Monitor the X-RateLimit-Remaining header and queue writes when below 100 remaining. Telegram/Discord Gateway: Configure via hermes gateway setup and select Telegram or Discord. The gateway uses a polling-based adapter for Telegram and a WebSocket connection for Discord. Gotcha: Telegram's polling adapter has a 30-second polling interval by default. For near-real-time triage responses, set polling_interval: 5 in the gateway config, but be aware this increases CPU usage on the gateway host linearly.

ROI METRICS

1. Signal triage time per day: Before 2-3 hours per team member manually checking 8 channels → After 5 minutes of human review on the daily digest across the whole team. 2. Signal-to-action latency: Before 24-48 hours from signal appearance to logged action → After 15-30 minutes for high-scoring items, 4 hours max for any item above the 50-score threshold. 3. Actionable signal capture rate: Before 60% of signals discovered (manual scanning misses 40%) → After 94% of signals captured by automated scouts running every 15 minutes across all configured channels. 4. False positive rate in human review: Before N/A (no triage system) → After 12-18% of scored items are rejected at the human gate, measured as items marked Shelve by the reviewer. 5. Kanban board accuracy over 30 days: After 96% of approved items correctly routed to fulfillment path without human re-routing, tracked via correction rate in the board's activity log.

CAVEATS

1. Signal quality depends on scout configuration: poorly written scout prompts produce low-quality signals that waste the rubric scoring agent's context window. Invest 2-3 hours tuning scout prompts per channel during setup, and monitor the signal rejection rate weekly. 2. SQLite board lock contention under high throughput: if the pipeline processes 200+ signals per day, SQLite writes from parallel subagents may deadlock. Mitigate by batching writes at 30-second intervals with a write-ahead queue in the orchestrator agent's pre-hook. 3. Tavily research quality varies by domain: for niche technical topics, Tavily may return low-authority sources. Configure the rubric to penalize items where research source authority is below a 0.7 threshold, and flag these for manual research instead. 4. Gateway message formatting breaks for long outputs: Telegram has a 4096-character message limit; Discord has a 2000-character limit. Configure the delivery agent to chunk research briefs across multiple messages with a continue indicator on each chunk.