Inspiration
Most coding agents hand you code they never ran, then forget every mistake when the session ends. I wanted an agent that proves its code works and gets better over time.
What it does
Give Forge a task in one line. It writes Python, runs it in an isolated sandbox, reads the real stdout/stderr, and fixes failures for up to 5 attempts. Every success and failure goes into a hash-chained memory ledger and is distilled into reusable skills, so the next similar task starts smarter.
How we built it
Forge runs one loop: Recall → Plan → Write → Run → Observe → Fix → Learn.
- Plan: ambiguous tasks are sent back with one clarifying question instead of being guessed at.
- Run: a sandbox with a temp workdir, scrubbed environment, no network, and a 30s timeout.
- Governance: deny-lists and AST checks block shell access, network calls, secrets, and repo writes before anything executes. A
HALT=1switch stops all runs instantly. - Anti-gaming: identical retries are detected as spinning and stopped, and empty output is rejected.
- Memory: a hash-chained ledger of episodes, distilled into skills.
Challenges we ran into
Making the agent trustworthy: stopping it from "passing" with empty output, and from looping on the same broken fix.
Accomplishments that we're proud of
57 offline tests, lint-clean, zero known dependency vulnerabilities, and an offline proof that anyone can run without an API key.
What we learned
What's next for Forge
Multi-language support and routing between Nemotron models (Nano, Super, Ultra) based on task difficulty.
Log in or sign up for Devpost to join the conversation.