The awesome expert Pablo Picasso is famously attributed alongside saying:
"It took me four years to coat akin Raphael, but a life to coat akin a child."
His work, Le Taureau (The Bull), captures this beautifully. Across 11 iterations, Picasso increasingly stripped distant the particulars of a masculine livestock until lone a few lines remained. Yet the masculine livestock was unmistakable.
In my journey of learning to build AI agents, I can relate. It has arrive complete circle.
Like many engineers, I started alongside abstractions and advanced application systems. Frameworks, orchestration engines, multi-agent architectures, province machines, vector databases, recollection systems, and detailed preparedness mechanisms.
The additional I experimented, the additional I established myself stripping things away.
The models were getting better. More things worked alongside a sole prompt. But that wasn't the entire story. I had been overcomplicating things.
This is now my workflow.
I open my terminal.
mkdir my_awesome_agent cd my_awesome_agent claude
and commencement a conversation.
"Claude, I desire to build an delegate that does XYZ. Help me initialize it. Create an AGENTS.md and let's fig out what alternatively it needs."
I activity through a few questions alongside Claude.

What is this delegate expected to accomplish?
What is the genuine job? Not what example to use, unless you're creating a framework. What does achievement appearance like? What decisions can the delegate create independently? What should it never do without permission?
What skills does it need?
Skills are specialized instructions that can be reused and remixed. There are plentifulness of outstanding examples on the internet. Borrow them, modify them, merge them, or ask Claude to create new ones.
Often, a accomplishment is a Markdown document alongside instructions and perchance a few supporting resources.
What environment does it need?
The delegate volition need context: what your endeavor does, who your customers are, how your codebase works, or several domain-specific knowledge. These can all live as Markdown records inner the folder. Some environment can be retrieved whenever needed. Some comes from the conversation itself. You don't automatically need a vector repository or a advanced recollection architecture.
What tools does it need?
To execute a job, an delegate needs tools. Maybe it needs to hunt the web, peruse a database, call an API, execute a script, or engage alongside a browser.
Introducing a tool is as uncomplicated as enabling a plugin or adding an MCP connector.
I ask Claude to assistance acknowledge the smallest set of tools required to accomplish the task. A tool could equal be created dynamically, alongside the delegate penning and invoking code. The delegate could additionally discover tools on the net and instal them.
How does the delegate get invoked?
For an delegate to be useful, it needs to be invoked as part of a workflow, whether by a cron job, a webhook event, or another program.
The simplicity is in the packaging. A directory alongside instructions is a spotless deployable unit. It can be versioned, copied around, and embedded in another processes.
With Claude and the required tools and credentials available, it can be invoked headless using the -p flag:
cd my_awesome_agent claude -p "Do notable stuff"
Headless method there is no interactive chat. Claude receives the task, plant in the agent's folder, prints the result, and exits. A manuscript can continue in the project and collect the output. For unattended work, I configure the tool permissions in advance. Claude can additionally come back organized JSON for another program to use.
Anthropic's documentation on headless execution covers the commands and configuration.
When the delegate starts, it needs to cognize anywhere to discover its instructions, what input it has received, and whether it should resume former work. The invocation scheme can alter without changing the essential intent of the agent.
This setup can be deployed on a distant VM, alongside hooks triggering the delegate and a tunnel providing admission to a assistance that invokes it.

How does it cognize whether it has done a fine job?
This is anywhere things have drastically improved. I can now get things done alongside a sole prompt, without detailed instructions. Nowadays, I frequently fair rely it to get on alongside the job. And I anticipate it to keep getting better.
When it gets item wrong, I go through it alongside Claude. What happened? What was missing? What should we change?
Claude saves local meeting transcripts under ~/.claude/projects/<project>/<session>.jsonl. Each copy records the conversation, tool calls, and tool results. It provides a trace of how the project unfolded: what Claude was asked to do, what it tried, what came rear from the tools, and anywhere errors appeared.
Claude can afterward be pointed at the copy and asked to analyze the run. Where did it go off track? Did it young female an instruction, deficiency context, or misinterpret a tool result? What evidence in the trace supports that explanation? That gives us item tangible to analyze before deciding what to change.
If the instruction is reusable, we preserve it in the instructions, context, examples, or tests.
Over time, the delegate improves. Not necessarily since its underlying example has changed, but since its operating environment and functioning instructions have rotate into better.
How do we cognize it is improving?
To be rigorous concerning the implementation, you need benchmarks: delegate tasks alongside apparent criteria for success. Running the identical tasks lets you difference your delegate throughout harnesses specified as Gemini CLI, Codex, or pi, or throughout models specified as Opus, Astra, Sol, and DeepSeek. It additionally tells you whether changes to your own setup are helping.
These tests are called evals. This is likely the hardest part of all of this. I volition portion my experiences alongside evals in a afterward post. In the meantime, these writeups explain them in much additional depth:
- Hugo Bowne-Anderson: Does Your AI Agent Actually Work? A Guide to Evals
- Hamel Husain and Shreya Shankar: AI Evals: Everything You Need to Know
Orchestration
A project can be damaged into smaller tasks and delegated to specialized subagents. In this setup, all subagent has its own directory under the chief agent, containing its instructions, context, and tools.
The base delegate acts as an orchestrator, or chief of staff. It decides which subagent to invoke, gives it a task, and brings the results together. For example, a chief-of-staff delegate could coordinate all kinds of action throughout a company, delegating activity to expert agents.

Making this concrete
I am using this setup to publish this website, including the part you are reading. For this project, I am using Codex.
The delegate plant in the website's project folder. AGENTS.md points it to my editorial direction and publishing instructions. Those records grasp my preferences and the steps for updating the site. Here is a shortened perspective of the genuine folder:
claramap/ ├── AGENTS.md ├── docs/ │ ├── editorial-memory.md │ └── publishing.md ├── content/posts/ │ └── what-picasso-taught-me-about-building-ai-agents.md ├── public/assets/blog/ └── scripts/ ├── build.mjs └── check-live.mjs
The delegate is connected to Cloudflare through MCP, so it can deploy the location as part of the workflow.
For this article, I asked it to add the orchestration division and create the chief-of-staff diagram. It edited the Markdown, generated the illustration, built the site, deployed to Cloudflare, and checked the live page.
The publishing instructions additionally shield Google Analytics and finding files. The build automatically updates the site's analytics tracking code for its current pages, regenerates llms.txt and llms-full.txt, and updates the RSS feed.
I assessment the outcome and accurate it in the identical conversation. Change the wording. Move a diagram. Make an depiction bigger. Reusable feedback goes rear into the editorial memory, so forthcoming runs have improved instructions to activity with.
The Essence
This is the foundation. Everything alternatively builds on this.
No need for OpenClaw, Hermes, Muse, LangChain, Pydantic, or any another delegate framework. Solve the essential case first, afterward add machinery later. You can discover leads, create content, publish, dispatch emails, answer questions, inhabit out forms, or compose code. Pick a job. Give it what it needs. Let it work.
Like Picasso's bull, my method became simpler as I learned what to depart out. The progression looks item akin this.
