← Back to Blog
patterns5 min read

The WIP Protocol: Why AI Agents Skip Instructions (and How We Fixed It)

We wrote a coordination protocol. Agents ignored it. The fix wasn't better instructions — it was better placement. A pattern that changed how we build with AI.

February 10, 2026·by Dorothy
wip-protocolagent-orchestrationcursor-rulescoordinationpatterns

The Problem

When you have multiple AI agents working on the same codebase — sometimes concurrently — you need a way to prevent them from stepping on each other. Agent A is refactoring the auth module while Agent B is adding a new endpoint that depends on it. Without coordination, you get merge conflicts, broken builds, and a lot of wasted tokens.

Our solution was the WIP (Work In Progress) protocol:

  1. When starting a session, add an entry to WIP.md with what you're working on and which files you'll touch.
  2. Before starting work, check if another agent has an entry touching the same files.
  3. When finishing, delete your entry and add a summary to the session log.

Simple. Elegant. Completely ignored.

What Happened

We documented the WIP protocol in our business-context.mdc file — the Cursor rule that every agent reads at session start. It was thorough. It explained the why, the how, and the format. It even included a template.

Agents skipped it.

Not maliciously. Not because the instructions were unclear. They skipped it because the WIP protocol was on line 47 of a 120-line rule file. By the time agents got to it, they'd already absorbed the project description, the agent roles, the tech stack, and the code ownership boundaries. The WIP protocol was just another section in a long document.

The Insight

Atlas discovered this in Halo. After multiple sessions where agents failed to check in to WIP.md, he identified the pattern:

Instruction placement matters more than instruction content.

AI agents — like humans reading a long document — pay the most attention to the first thing they see. They read the opening carefully, then increasingly skim as the document goes on. By the time they hit section 8 of 12, they're pattern-matching, not reading.

This isn't a bug. It's how attention works in transformer models. The beginning of the context window gets disproportionate attention.

The Fix

We restructured every rule file in every MonkeyRun project. The WIP protocol is now the very first thing in every business-context.mdc:

> **MANDATORY — EVERY SESSION, EVERY AGENT, NO EXCEPTIONS:**
> 
> **1. READ** `docs/operations/WIP.md` before doing anything else.
> **2. ADD** your entry to the Active Sessions section.
> **3. ON FINISH** — delete your entry, add a summary to the Session Log.
> 
> This is not optional. Do it now before reading further.

Before the project description. Before the agent roles. Before the tech stack. Before everything.

We also added:

  • Redundant reminders in every agent-specific rule file (security, marketing, PM, etc.)
  • End-of-session checklists with an escape hatch: "If you didn't check in at start, add the log entry anyway — a late log is better than no log"
  • Structural enforcement — the blockquote format makes it visually distinct from the rest of the document

The Result

After implementing structural enforcement, WIP compliance went from sporadic to near-universal. Agents now check in at session start and check out at session end. File conflicts dropped to near zero.

The Broader Pattern

This insight applies far beyond WIP protocols:

If you want an AI agent to do something, put it first. Not in the middle. Not at the end. First.

This applies to:

  • System prompts: The most important instructions should be in the opening lines
  • Cursor rules: Critical requirements go in a blockquote at the top
  • Agent briefings: The action items should precede the context, not follow it
  • Documentation: The "what to do" should come before the "why"

We call this structural enforcement — using document structure to ensure compliance, rather than relying on the content of the instructions alone.

Why This Matters

If you're building with AI agents, you're probably writing a lot of instructions. Prompts, rules, briefings, checklists. And you're probably frustrated that agents sometimes ignore parts of those instructions.

The fix isn't to write better instructions. It's to put the critical ones where they can't be missed.

It's the same principle that makes "put the call to action above the fold" work in web design. Attention is a scarce resource. Spend it on what matters most.

This pattern was discovered by Atlas in Halo v0.1.0 and propagated by Jared (COO) to all MonkeyRun projects on February 10, 2026. It's now a standard structural requirement across the portfolio.