DEV Community

BlockAIx
BlockAIx

Posted on

🚀 An Autonomous Development Runner That Executes a ROADMAP.json

#agents #ai #automation #showdev

Agentik

For the past few months, I’ve been testing one idea:

What if software projects were executed from a structured roadmap — automatically?

Not random file generation. Not chatbot-style commits.

I wanted a controlled loop: build → test → fix → static checks → docs, driven by one source of truth.

So I built Agentik.

Agentik is an open-source autonomous development runner that executes every task in ROADMAP.json through a dependency-aware pipeline.

Why this exists

Most AI coding tools generate code, but don’t manage execution.
CI manages execution, but doesn’t generate code.

Agentik sits in the middle:

  • Reads a task graph
  • Runs tasks in dependency order
  • Tests and auto-fixes failures
  • Tracks token/cost usage
  • Supports parallel-safe execution
  • Resumes after interruption
  • Can manage git history when git automation is enabled

What it runs

Define tasks in ROADMAP.json (or do it with a model using the AGENTS.md), then run:

python agentik.py
Enter fullscreen mode Exit fullscreen mode

Pipeline:

  • Build — implement code + tests
  • Deps — install declared dependencies
  • Test — run ecosystem tests
  • Fix — patch failing tests (bounded retries)
  • Static — lint/type checks
  • Static Fix — patch static failures (bounded retries)
  • Doc — update README
  • Commit — when git management is enabled
  • Deploy — optional hook if configured

The key insight

AI performs far better with structured work:

  • Atomic tasks
  • Explicit dependencies
  • Output-scoped files
  • Clear acceptance criteria

Instead of “build a SaaS app,” you define deterministic units and execute them reliably.

Built-in guardrails

  • Single root task + dependency validation
  • Parallel safety via disjoint outputs
  • Saved state after phases (resume on Ctrl-C)
  • Budget limits + estimated USD tracking

Autonomy without rails is chaos. Agentik is about controllable autonomy.

Supported ecosystems

  • Python (pytest + ruff)
  • Node/TypeScript (vitest/jest + tsc)
  • Deno
  • Go
  • Rust

You can also define your own ecosystem manually.

Try it

  1. Install opencode (and Python :))
  2. Clone the repo
  3. Create projects/<name>/ROADMAP.json
  4. Run python agentik.py

GitHub: https://github.com/BlockAIx/agentik

If you try it, I’d love feedback on failure handling, scheduling, and real-world task design patterns.

Author: Thomas Heymelot (BlockAIx)
LinkedIn

Top comments (0)