v0.9.3 · an ongoing exploration

Understanding agents by building with them.

I wanted to understand how agent harnesses work, and why they work the way they do. So I started building one.

Since April 2026A personal project that started as a question and kept growing.
Claude builds, ChatGPT reviewsI am not a developer. Claude writes most of the code, ChatGPT reviews it, and I decide what goes in.
One user so farI use it for my own tasks. Nothing here has been tried at scale.

zsh

$ npm install -g nodal-agents

$ nodal-agents up

What it looks like today

The Nodal-Agents dashboard in the light theme, showing agent, skill and connector counts, job totals, weekly activity and a per-agent token table. — The dashboard, light theme.

An agent detail page in the dark theme, showing its model, role, skill and connector counts, success rate, a chart of its runs over seven days, and its attached skills. — One agent, dark theme.

01The starting point

A question that became a working harness

It started with questions about the agent tools I was already using.

Quentin Beau · User Experience Expert · Exploring agentic systems through practice

Working on agentic systems, I wanted to get closer to the decisions behind the tools we use. How do agents share work? What gives them permission to act? What happens when something goes wrong?

In April 2026, I started building a small harness to explore those questions, with Claude as the main developer and ChatGPT as a reviewer. I am not a developer, though I have a solid understanding of software development methods.

The project kept growing. Today, I use Nodal for some everyday tasks, to try different models, and to test other ways of coordinating agents. It gives me somewhere to change an approach and see how it behaves in practice.

02The questions

What I am trying to understand

The harness lets me experiment with the mechanics behind an agent’s behaviour. These are the questions I keep coming back to.

How should agents share the work?

The question: When does it help to hand a task to a specialist? When does splitting it across several agents add more coordination than value?
What I can test in Nodal: Route a request to one agent, or distribute work across several. Follow the child tasks and how their results return to the orchestrator.

Where should autonomy begin and end?

The question: Which decisions can an agent take on its own? When is an approval useful, and what information helps someone make that decision?
What I can test in Nodal: Set autonomy and tool access per agent. Try approval steps that explain the intended action and its impact, with limits on execution.

What do I need to see to follow the work?

The question: How can I tell what happened, what changed, and what still needs checking? What should be visible when an agent reports that it has finished?
What I can test in Nodal: Inspect run histories, changed files, verification output and reviewer findings as separate records. Trace errors through delegated tasks.

What changes when the model or context changes?

The question: How much of an outcome comes from the model, its instructions, the tools it can reach, or the information it remembers?
What I can test in Nodal: Choose different models per agent, assign skills and connectors, and work with persistent memory across tasks.

These are areas of exploration, not results. My own use gives me observations to discuss and approaches to keep testing.

03The harness today

A place to put those questions to work

Nodal runs on your machine and connects to the models and tools you choose. This is what the working project looks like today.

Different agents, different setups

Each agent can have its own model, tools, skills and memory settings. Tasks can be delegated to other agents.

Real tasks and connected tools

Try research, document work or repository tasks through the dashboard, messaging channels, schedules and webhooks.

Visibility and ways to intervene

Inspect execution records, configure approvals and limits, and use workspace snapshots to recover earlier file states.

Models, connectors and other building blocks

Models: Hosted providers including Anthropic, OpenAI, Google and OpenRouter, plus local models through tools such as Ollama and LM Studio.
Tools and context: Connectors for services such as Gmail, Notion and Google Drive. MCP servers extend access to other tools. Skills provide written guidance.
Ways to interact: The dashboard, Telegram, Discord, Slack and WhatsApp. Tasks can also start from a schedule or a webhook.

04Building and checking

The experiment includes how I build it

Claude handles the main development work, with ChatGPT as a reviewer. I use this setup to turn questions into changes I can try in the harness.

The repository also includes automated checks, tests and a public quality report. They make the state of the project easier to inspect as it evolves. These figures are measured, not estimated.

34packages measured

10,054test cases

846test files

298end-to-end cases

80.8%line coverage

22 / 24capabilities green at both levels

Measured by the nightly run on 29 September 2026, commit 13f5a066. See the run. 22 of the 24 capabilities are green at both levels today. The other 2 are missing one level, and the page that tracks them says so in grey rather than in green.

Keep changes inspectable
Development happens through small pull requests, with automated checks and review.
Check what actually happened
The harness keeps file changes, verification output and reviewer findings available to inspect.
Make room for another perspective
Sharing the project is a way to get feedback on the choices, discover different use cases and compare approaches.

Pre-1.0

The quality report

Nodal is still pre-1.0. I use it for personal tasks while continuing to explore its behaviour and limits. Tests and reviews are part of that process, and the report shows them as they are, including what is not measured.

See the quality report

Where it stands: A working project used for personal tasks and ongoing experiments. It continues to change, and updates can include breaking changes.
Where the exploration continues: Different models, new use cases, and ways to make delegation, autonomy and control work together. Other people’s experiences are part of what I want to learn from next.

0.9.3latest on npm · September 2026

Read the changelog

05Try it and compare notes

Curious? Take a closer look

You can run Nodal locally and explore the current implementation. The documentation covers setup and configuration.

Requires Node.js 22 or later. Connect your own model provider or a local model. Provider usage may incur costs.

Read the setup guide nodal-agents on npm

Compare notes

I’d like to hear what you find

Maybe you’ve tried another way of handling approvals. Maybe delegation worked differently from what you expected. Or perhaps there’s a task you’d like to explore with it.

That’s why I’m sharing the project: to get other perspectives on something I’ve mostly been exploring through my own use. An issue is a good place for a concrete observation, a question or an idea to explore.

Share feedback on GitHub