Back to browse
GitHub Repository

Hatice is an autonomous coding agent orchestration system.

141 starsTypeScript

Hatice – Autonomous Issue Orchestration with Claude Code Agent SDK

by mksglu2·Mar 7, 2026·1 point·0 comments

AI Analysis

●●●BangerBig BrainBold BetWizardry

Claude agents solving GitHub/Linear issues autonomously—production-grade, 4K LOC tested, real demo.

Strengths
  • Real orchestration runtime—4,148 LOC + 328 tests, zero type errors, not a toy or wrapper
  • Specification-driven (single WORKFLOW.md + .claude/CLAUDE.md) makes intent explicit and reproducible
  • MCP tools for agents to query trackers directly, cost tracking per session, exponential backoff retry
Weaknesses
  • Harness Engineering manifesto is aspirational; unclear if agents reliably solve real production issues
  • Depends on Claude Code agent SDK; locked to Anthropic's tooling, limited to Linear/GitHub
Target Audience

Engineering teams testing agent-driven development, autonomous coding workflows

Similar To

Symphony (OpenAI/Codex) · LangGraph agent loops · Devin (Cognition)

Post Description

Hatice polls your issue tracker (Linear or GitHub Issues), spins up isolated workspaces, and dispatches Claude Code agents to solve each issue end-to-end — dispatch, multi-turn execution, retry with exponential backoff, reconciliation, and real-time observability.

Inspired by OpenAI's "Harness Engineering" manifesto and Symphony (Elixir/OTP), rebuilt from scratch in TypeScript with the Claude Code Agent SDK.

Key differences from Symphony: - GitHub Issues support (not just Linear) - SSE real-time dashboard (no WebSocket dependency) - Per-session cost tracking in USD - Fine-grained tool control (allow/disallow specific tools) - MCP server tools for agents to query Linear/GitHub APIs directly - Single WORKFLOW.md file configures everything

4,148 lines of source, 328 tests, zero type errors. Every line written by AI agents.

Demo video in the README shows hatice building a presentation website autonomously — 5 Linear issues dispatched in parallel, each agent working in isolated workspaces with live hot-reload.

Similar Projects