Hatice: I Stopped Writing Code. Agents Do It Now.

5 Linear issues. 5 Claude agents. Zero lines written by a human. Here's how.

The Idea

OpenAI published "Harness Engineering" in February 2026. The thesis: engineers don't write code anymore. They design environments, specify intent, and build feedback loops. Agents do the rest.

Codex proved the concept. Symphony showed the architecture in Elixir/OTP. I wanted the same thing with Claude Code.

So I built Hatice.

What It Does

Hatice polls your issue tracker — Linear or GitHub Issues. When it finds assigned work, it spins up an isolated workspace, clones your repo, and dispatches a Claude Code agent to solve the issue end-to-end. Tests first, implementation second, commit when done.

Not one agent. Multiple. In parallel. Each in its own workspace. Each with its own session. Each tracked independently.

The full lifecycle: dispatch, multi-turn execution, retry with exponential backoff, reconciliation, workspace cleanup. If an agent crashes, the supervisor restarts it. If the API rate-limits you, it backs off and retries. If the issue gets closed while the agent is working, it detects the state change and cleans up.

One config file controls everything:

tracker:
  kind: linear
  apiKey: $LINEAR_API_KEY
  projectSlug: "TEAM"
  assignee: "me"
workspace:
  rootDir: /tmp/workspaces
hooks:
  afterCreate: "git clone repo . && npm i"
claude:
  permissionMode: bypassPermissions
  model: claude-sonnet-4-20250514

That's it. WORKFLOW.md. YAML frontmatter for config, markdown body for the agent prompt. LiquidJS templates inject issue details: {{ issue.title }}, {{ issue.description }}.

The Demo

I pointed Hatice at a blank HTML file. Created 5 Linear issues — hero section, features grid, how-it-works flow, configuration showcase, footer. Moved them all to "In Progress" at once.

Five agents spun up simultaneously. Each picked a section. Each read the design system from .claude/CLAUDE.md. Each wrote code to the same file through symlinked workspaces. Live-server hot-reloaded every change.

The presentation website for Hatice was built by Hatice. While I watched.

What I Learned Building This

Strip the CLAUDECODE env var. Claude Code sets this to prevent nested sessions. When you're spawning agents from inside a Claude session, you have to delete it from the child's environment. Cost me hours of debugging agents that exited immediately with code 1.

allowDangerouslySkipPermissions is required. If you set permissionMode: 'bypassPermissions' in the SDK, you also need this flag set to true. Otherwise agents complete in one turn without doing anything. The SDK silently accepts the config but doesn't actually bypass.

maxTurns: 10 is too low. Agents spend their first 5-6 turns exploring the codebase. Set it to 0 (unlimited) and let the agent decide when it's done.

Polling interval matters. Default 30 seconds feels like forever when you're demoing. 5 seconds is responsive enough without hammering the API.

Symphony vs Hatice

Symphony is Elixir/OTP — BEAM VM, Phoenix LiveView, GenServer supervision trees. Beautiful architecture. Wrong stack for my team.

Hatice is TypeScript — Node.js, Hono for HTTP, Pino for logging, Zod for validation. Same architectural patterns, different runtime. Claude Code Agent SDK instead of Codex JSON-RPC.

What Hatice adds: GitHub Issues support, SSE dashboard (no WebSocket dependency), per-session cost tracking in USD, fine-grained tool control, MCP server tools so agents can query your tracker's API directly.

4,148 lines of source. 328 tests across 30 files. Zero type errors. Every line written by AI agents, guided by humans.

Should You Use This

If you're already using Claude Code and have a structured tracker workflow — yes. Point it at your backlog. Start with low-risk issues. Watch how agents handle them.

If you're looking for magic — no. Agents still need good specifications. Vague issues produce vague code. The quality of your issue descriptions is the quality of your output.

Hatice doesn't replace engineering judgment. It replaces the typing.

Open source. MIT. github.com/mksglu/hatice

Mert Koseoglu, Senior Software Engineer, AI consultant. x.com/mksglu · linkedin.com/in/mksglu · mksg.lu