Noureddine RAMDI / Aeon: a TypeScript Autonomous Agent Framework with a self-healing skill architecture

Created Tue, 05 May 2026 13:37:39 +0000 Modified Sat, 23 May 2026 20:41:27 +0000
854 Words
github-stars typescript autonomous-agent github-actions self-healing ai-framework

aaronjmars/aeon

Aeon is a TypeScript framework that rethinks autonomous AI agents by running them as scheduled GitHub Actions workflows without manual approvals. Its standout feature is a closed-loop self-healing architecture that automatically detects, evaluates, and repairs failing skills, enabling agents to run unattended and improve over time.

What Aeon does and how it works

Under the hood, Aeon is built as an autonomous agent framework where agents execute skills on a schedule, typically three times daily, entirely within GitHub Actions. This zero-infrastructure deployment means you don’t need to manage servers or containers — the framework leverages GitHub’s CI/CD environment as the runtime.

The codebase is TypeScript-based, with a modular architecture around a skill system currently offering over 90 skills. These skills cover domains like research, development, cryptocurrency, social media, and productivity, letting you compose agents with diverse capabilities. Skills are independently installable and manageable, supporting flexible customization.

A key architectural concept is persistent memory across runs, allowing agents to maintain context and state without external databases. Aeon also supports spawning fleets of agent instances specialized for particular tasks, improving scalability and separation of concerns.

Authentication integrates with Claude OAuth tokens or Anthropic API keys, and optionally uses the Bankr Gateway to optimize API usage costs. This shows a practical approach to managing AI service credentials and controlling expenses.

The self-healing skill loop and quality scoring

What sets Aeon apart is its automated maintenance loop for skills. Each agent run produces a quality score from 1 to 5 using a Haiku-based metric. This score is tracked over a rolling 30-run history to monitor skill health.

The self-healing loop consists of several phases:

  • Heartbeat: Runs three times daily to trigger agent execution and monitor overall system status.
  • Skill-health audits: Review the quality scores and usage metrics to detect underperforming skills.
  • Skill-evals: Execute assertion tests against skills to verify correctness.
  • Skill-repair: Automatically patches or disables skills that fail the evaluations.
  • Self-improve: Evolves prompts or configurations based on feedback from previous runs.

This closed-loop autonomy reduces the need for human intervention, which is often a bottleneck in autonomous agent deployments. It’s a technical tradeoff, as this complexity in automation requires careful design and robust testing, but it can significantly improve reliability and resilience in production.

The persistent memory and rolling quality scores add a feedback mechanism that enables agents to adapt and self-correct over time, making the system more robust in the face of external API changes or skill regressions.

Quick start

Getting Aeon running is straightforward if you have a GitHub account and basic familiarity with CLI:

git clone https://github.com/aaronjmars/aeon
cd aeon && ./aeon

This launches a local dashboard on http://localhost:5555 where you can:

  1. Authenticate with your Claude API key or OAuth token.
  2. Add communication channels like Telegram, Discord, or Slack for interaction.
  3. Select and schedule skills, optionally setting variables to focus their behavior.
  4. Push your configuration to GitHub, triggering Actions to take over execution.
  5. Run ./onboard to verify that all secrets, workflows, and notifications are properly set up. Append --remote to run the check inside GitHub Actions and receive results in your notification channel.

To add external skills from other repositories or Aeon’s own catalog, use the ./add-skill command with arguments to list, install specific skills, or install all:

./add-skill BankrBot/skills --list          # browse external repo skills
./add-skill BankrBot/skills bankr hydrex   # install specific
./add-skill BankrBot/skills --all           # install everything

./add-skill aaronjmars/aeon --list          # browse Aeon catalog
./add-skill aaronjmars/aeon token-alert monitor-polymarket  # install specific
./add-skill aaronjmars/aeon --all            # install everything

Installed skills appear in the skills/ directory and are disabled by default in aeon.yml. Flipping enabled: true activates them.

verdict

Aeon offers a compelling framework for developers interested in autonomous agents that can run fully unattended with a focus on resilience and self-maintenance. Its use of GitHub Actions for execution is clever — it removes infrastructure overhead but ties you to the GitHub ecosystem and its scheduling constraints.

The self-healing skill loop is the most interesting technical pattern here, providing a feedback-driven autonomy uncommon in open agent frameworks. However, this complexity means Aeon may not be a plug-and-play solution; users should be comfortable with TypeScript, GitHub workflows, and some manual configuration.

Its modular skill system and persistent memory make it well-suited for experimentation across domains like research and productivity, but those seeking more turnkey or cloud-agnostic agents might find it limiting.

All told, Aeon is worth exploring if you want to build autonomous agents that manage their own health and run with minimal ongoing human supervision, especially if you already use GitHub Actions as part of your toolchain.

→ GitHub Repo: aaronjmars/aeon ⭐ 270 · TypeScript