Workflow Agent

An autonomous enterprise workflow system that monitors GitHub and Slack in real time, uses LLMs to classify and analyse events, coordinates specialist AI agents, and delivers structured reports — all without human intervention.

---

What it does

When a GitHub issue or pull request is opened, the system automatically:

1. Receives the event via webhook
2. Classifies it using an LLM orchestrator
3. Spawns specialist sub-agents to analyse, gather context, communicate, and document
4. Posts a structured analysis to Slack
5. Comments directly on the GitHub issue or PR
6. Writes an immutable audit trail
7. Updates a live dashboard in real time

Zero manual intervention. Every step is logged, traced, and observable.

---

Architecture

flowchart LR
  %% Entry points
  gh[GitHub Webhooks]
  sl[Slack Events]

  %% Ingestion and orchestration
  api[[Express Webhook Server]]
  q[(pg-boss Queue\nPostgreSQL-backed)]
  orch{{Orchestrator LLM\nGroq + workflow plan}}

  %% Specialist agents
  subgraph agents[Specialist Agent Pod]
    direction LR
    data[Data\ncontext extraction]
    analysis[Analysis\nseverity + root cause]
    comms[Comms\nSlack + GitHub response]
    docs[Docs\naudit trail + completion]
  end

  %% Persistence and observability
  db[(Neon PostgreSQL\nevents · workflows · steps · llm_logs · audit_log)]
  dash[[Next.js Dashboard\nVercel + live SSE]]

  gh --> api
  sl --> api
  api --> q
  q --> orch

  orch -->|critical_bug_triage / pr_review_assist| data
  orch -->|slack_reply / general| analysis
  data --> analysis --> comms --> docs
  docs --> db

  data -. context .-> db
  analysis -. token log .-> db
  comms -. messages .-> db
  dash <-->|stream + detail views| db

  classDef source fill:#0f172a,stroke:#38bdf8,color:#e2e8f0,stroke-width:1.5px;
  classDef core fill:#111827,stroke:#f59e0b,color:#f9fafb,stroke-width:1.5px;
  classDef agent fill:#1f2937,stroke:#34d399,color:#f9fafb,stroke-width:1.5px;
  classDef store fill:#052e16,stroke:#22c55e,color:#ecfdf5,stroke-width:1.5px;
  classDef ui fill:#312e81,stroke:#a78bfa,color:#f5f3ff,stroke-width:1.5px;

  class gh,sl source;
  class api,q,orch core;
  class data,analysis,comms,docs agent;
  class db store;
  class dash ui;

Loading

---

Live demo

• Dashboard: https://workflow-agent-virid-one.vercel.app
• Backend: https://workflow-agent-backend.onrender.com

---

Key features

Multi-agent orchestration — A central orchestrator LLM classifies each event and delegates to four specialist agents, each with a narrow system prompt and a single responsibility.

Durable job queue — Uses pg-boss, a PostgreSQL-backed job queue. No Redis. No extra services. Survives server restarts with automatic retries.

GitHub issue + PR handling — Analyses both issues and pull requests. Posts risk assessments and fix recommendations as comments directly on GitHub.

Slack notifications — Structured, severity-coded messages posted to a Slack channel on every workflow run.

Immutable audit trail — Every action taken is logged: what happened, which agents ran, what they decided, how long it took, how many tokens were used.

Token cost observability — Every LLM call is logged with input tokens, output tokens, latency, and cost. Aggregated totals visible on the dashboard.

Live dashboard — Real-time Server-Sent Events (SSE) stream workflow updates to the Next.js frontend without polling.

Human-in-the-loop ready — Orchestrator can flag workflows requiring human approval before agents execute.

---

Tech stack

Layer	Technology
Backend	Node.js + Express
Queue	pg-boss (PostgreSQL-backed)
Database	Neon PostgreSQL
ORM	Prisma
LLM	Groq API (Llama 3.3 70B)
Slack	Slack Web API
GitHub	GitHub Webhooks + REST API
Frontend	Next.js + Tailwind CSS
Backend hosting	Render
Frontend hosting	Vercel

---

Database schema

Five tables power the entire system:

• events — normalised incoming webhooks with source, type, payload, status
• workflows — one row per workflow run, stores the LLM plan and current status
• workflow_steps — one row per agent execution, stores input and output
• llm_logs — every LLM call with model, tokens, latency, and cost
• audit_log — human-readable summary written by the Docs agent after each workflow

---

Workflow templates

The orchestrator picks from four templates based on the event:

Template	Trigger	Agents
critical_bug_triage	Critical bug issues	data → analysis → comms → docs
pr_review_assist	Pull request opened	data → analysis → comms → docs
slack_reply	Slack mention	analysis → comms → docs
general	Any other event	data → analysis → comms → docs

---

Agent responsibilities

Data agent — Extracts structured context from the raw event payload. For issues: title, body, labels, repo. For PRs: branch names, file count, additions, deletions.

Analysis agent — Sends context to Groq LLM with a focused prompt. Returns severity (low/medium/high/critical), likely cause, recommended action, and a human-readable summary.

Comms agent — Posts a formatted Slack message with severity emoji and full analysis. Posts a comment on the GitHub issue or PR via the GitHub REST API.

Docs agent — Writes a structured audit log entry, marks the workflow as completed, and updates the event status.

---