InspirationProject name:
AgentForge — Hire an AI Team With One Sentence
Elevator pitch: Describe a business goal; a team of specialized AI agents plans, researches, analyzes, and delivers ranked recommendations.
Track: AI + Business
Project story: Problem: Small businesses and solo operators sit on raw business data — churn exports, sales CSVs, support tickets — but turning that data into decisions requires an analyst team they can't afford. Single-shot AI chatbots summarize; they don't investigate. They can't plan a multi-step analysis, run code on data, check their own work, or hand over a decision-ready report.
Solution: AgentForge takes one plain-language business goal ("Analyze why our churn spiked in Q3 and recommend 3 fixes") and spawns a team of specialized AI agents — Planner, Researcher, Analyst, Forecaster, Critic — that work the goal, stream progress live, and deliver a structured report: key insights, charts, predictions, and ranked recommendations. Three prebuilt team templates (Market Research, Sales Insights, Support Triage) cover the most common jobs, and one click exports any finished run as an n8n workflow.
Technical approach: The Planner breaks the mission into a DAG of subtasks with agent assignments; workers execute with retries; a Critic agent reviews every output and flags weak claims before the report is finalized. Analyst agents run real Python on uploaded CSVs (stats, correlations, trend detection, chart specs) inside a restricted sandbox — not LLM guesswork. Provider-agnostic LLM layer, realtime run view over websockets, per-agent token/cost meter on every run. Frontend: Next.js 14. Backend: FastAPI + SQLAlchemy. Every run exportable as Markdown/PDF, plus one-click n8n workflow JSON.
Built with: Next.js 14, FastAPI, Python, TypeScript, Tailwind CSS, WebSockets, SQLAlchemy, pandas, matplotlib, n8n
What it does
AgentForge lets you hire an AI team with one sentence. You type a plain-language business goal — "analyze why our churn spiked in Q3 and recommend 3 fixes" — pick a prebuilt team template (Market Research, Sales Insights, Support Triage) or let the Planner assemble one, upload your data, and hit run. A team of specialized agents — Planner, Researcher, Analyst, Forecaster, Critic — then works the mission live: the Planner breaks it into subtasks, the Researcher gathers context, the Analyst runs real Python on your CSVs (stats, correlations, trends — not guesses), the Forecaster predicts what's next, and the Critic reviews every claim before it ships. You watch it all stream in real time, with every agent's token cost metered live. The deliverable is a structured report — ranked insights, charts, predictions, ranked recommendations — downloadable as Markdown or PDF. And one click exports the entire run as an n8n workflow, turning the recommendation into automation.
How we built it
Frontend in Next.js 14 (App Router, Tailwind), backend in FastAPI + SQLAlchemy, with websockets streaming live run state to the browser. The agent framework is the core: the Planner decomposes goals into a DAG of subtasks, workers execute with retries, and the Critic does a full review pass before anything reaches the report. Analyst agents execute real Python inside a restricted sandbox (pandas, matplotlib for chart specs). The LLM layer is provider-agnostic — any OpenAI-compatible endpoint works. ~37,000 lines, 42 test files, compile-clean.
Challenges we ran into
Orchestration reliability was the big one — getting five agents to hand off work through a DAG without dropping context or looping forever took real iteration on the planner and retry logic. Sandboxing the Python execution safely while still letting analysts do useful statistics was a careful balance. And streaming coherent live state over websockets (agent timeline, artifacts landing, cost meter ticking) without the UI desyncing mid-run was harder than it looks.
Accomplishments that we're proud of
A genuinely working multi-agent team from a single sentence — not a chatbot wrapper. The Analyst runs real code on real data, which is the difference between "AI demo" and "analyst replacement." The per-agent token/cost meter makes AI spend transparent mid-run. The n8n export is our favorite: insight becomes automation in one click. And the whole thing is tested — 42 test files, not vibes.
What we learned
The Critic agent is what separates a toy from a tool — without a verification pass, agents hallucinate with total confidence, and no business user can trust the output. We also learned that visible cost metering changes behavior: when users see tokens ticking, they write better goals. And that demo mode (deterministic, offline) is worth building alongside live mode — it's the only way to demo reliably.
What's next for AgentForge
We're already building it: real data connectors (Postgres, BigQuery, Snowflake, Google Sheets, Stripe, Shopify) replacing CSV upload, scheduled recurring missions with diff reports and alerts, new agent types (SQL data-engineer, competitor intel, quant forecaster with real statistical models), persistent dashboards with shareable links, multi-tenancy with proper logins and roles, and an evaluation harness that regression-tests agent outputs. The codebase is already past 60k lines on the way there.
Built With
- fastapi
- matplotlib
- next.js-14
- pandas
- python
- sqlalchemy
- tailwind-css
- typescript
- websockets
Log in or sign up for Devpost to join the conversation.