Some CLAUDE.md or AGENTS.md files I've seen are micromanagement dressed up as engineering. Hundreds of lines dictating which files to read, which commands to run, what order to do things in. Some hit 1000s of tokens before the agent starts thinking about your actual task.
Every token spent on instructions is a token the agent can't spend on your work. And the more rigid the instructions, the more brittle the agent becomes.
I developed a 28-line operating law for autonomous agents — born from wanting to give the agent a task in Claude Code terminal and have it complete it without any input or feedback, no matter how easy or complex. Here's what I discovered.
The plan mode trap
I asked the agent to build a full-stack app. Set it loose with --dangerously-skip-permissions and waited. 34 minutes later: zero files created.
The agent had entered plan mode — voluntarily switching itself to read-only. It spawned three research subagents, produced excellent analysis, then sat there. Planning. Unable to write a single file.
Turns out plan mode isn't a hard system constraint. It's a prompt injection — a system reminder prepended to every message saying "you MUST NOT make any edits." The write tools are still there. The agent just chooses not to use them.
The fix was one line:
Never use EnterPlanMode. Execute tasks directly.
EnterPlanMode isn't even a real tool name. But the agent understood the intent and stayed in execution mode.
The memory problem
Agents have no memory between sessions. Every new session starts fresh — no idea what was decided or what's half-built. Anthropic's own research confirms this as a core failure mode.
Two more lines fixed it:
You have no lasting memory between sessions. Use .memory-bank in project root for persistence.
Read .memory-bank at start, write to it throughout. Goal, success measures, plan, and progress must exist there before and during execution.
The agent now knows it has no memory, knows where to persist, and knows what to persist. How it organises those files is its decision. This is a deliberate departure from memory bank patterns like those in Cline and Roo Code, which prescribe exact file structures, naming conventions, and update protocols. I tell the agent what to remember, not how to store it. A capable model can structure its own memory.
Before and after
Same task, same model, same permissions. Only the CLAUDE.md changed.
Before: 34 minutes in plan mode. Zero files. Stuck in read-only research. Three subagents doing excellent work — but nothing written to disk.
After adding Never use EnterPlanMode. Execute tasks directly.: The agent stayed in execution mode and started building immediately.
After adding the .memory-bank lines: The agent's first action was reading .memory-bank, finding a clean slate, writing goal and success measures, writing an architecture plan — then building database, API adapters, routes, server, and full frontend. All within minutes.
Delegation, not control
The deeper shift isn't about CLAUDE.md syntax. It's about how we work with agents.
We carry habits from tools that needed precise instructions. Compilers need exact syntax. Pipelines need exact step sequences. We learned that computers do exactly what you tell them, nothing more.
Frontier models reason. They infer. They make judgment calls. The more you constrain their judgment with rigid instructions, the worse those calls become — you've filled their context with rules instead of understanding.
The shift is from control to delegation. From "do steps 1 through 47" to "here's what success looks like, here are the boundaries, go."
If you have to specify everything, you haven't delegated. You've written a very slow script.
The full Autonomous Agent Law and reasoning behind each line is on GitHub. Fork it, test it, see what changes.
The agent is capable. Stop writing it a checklist. Write it a constitution.
Dawid M. — Consultant. 20 years of strategy, systems, and data across manufacturing, finance, real estate, and telco. Working on making AI agents deliver business value.
Top comments (0)