← All projects

Floor Agents #

In DevelopmentUpdated Apr 9, 2026

Your AI dev team. Plugs into your tools. Ships code while you sleep.

Overview #

Floor Agents is an autonomous AI engineering platform that gives software teams specialized AI agents capable of decomposing tasks, writing code, creating pull requests, reviewing code, and writing tests. It integrates with your existing tools — GitHub, Linear, and any LLM provider.

How It Works #

A configurable team of AI agents collaborates through a structured workflow:

Agent Role
PM Decomposes tasks into sub-issues
Backend Dev Writes backend code, creates PRs
Frontend Dev Writes frontend code, creates PRs
CTO Reviews PRs, approves or requests changes
QA Writes tests, validates functionality

The workflow follows a state machine:

Backlog → Triage → In Progress → In Review → QA → Done
                        ↑               │
                        └── Changes Requested (max 3 cycles → escalate to human)

Every PR is reviewed by the CTO agent. If changes are needed, the dev agent revises — up to 3 cycles before escalating to a human. All work is visible in your issue tracker and PRs. No black box.

Tool Agnostic #

Floor Agents connects to the tools you already use:

  • Git platforms — GitHub (GitLab, Bitbucket planned)
  • Task managers — Linear, GitHub Issues (Jira planned)
  • LLM providers — Claude (Anthropic), OpenAI, Google Gemini, LM Studio, Ollama, Groq, Together

Use cheap local models for routine dev work, expensive cloud models for critical reviews. Mix and match per agent.

Architecture #

Built as a Bun monorepo with 10 packages:

  • @floor-agents/core — Types, config loader, YAML validation
  • @floor-agents/orchestrator — Main loop, state machine, guardrails, cost tracking
  • @floor-agents/context-builder — File selection via import tracing, prompt rendering, token budgets
  • @floor-agents/anthropic — Claude API adapter
  • @floor-agents/claude-code — Claude Code CLI adapter (native worktree execution)
  • @floor-agents/openai — OpenAI-compatible adapter
  • @floor-agents/gemini — Google Gemini adapter
  • @floor-agents/lmstudio — Local model adapter
  • @floor-agents/github — GitHub REST API (branches, commits, PRs)
  • @floor-agents/task — Task manager adapters

Key Design Decisions #

  • Config-driven — A single YAML file defines team composition, models, guardrails, cost limits, and workflow states
  • Native worktree execution — Dev agents spawn Claude Code CLI directly in git worktrees with full file access, not just tool-use
  • Crash-recoverable — 10-step execution state machine with file-based persistence. Every step is idempotent
  • Cost controls — Per-task and per-day spending limits. Local models report $0
  • Guardrails — File count/size limits, blocked paths (.env, .pem, CI configs), path traversal detection
  • Pluggable adapters — Task managers, git platforms, and LLM providers all behind interfaces. Add new integrations without touching core logic

Tech Stack #
Language TypeScript (strict)
Runtime Bun
Config YAML
State File-based JSON (PostgreSQL planned)
Tests 116 tests across 22 files