How Do Senior Engineers Steer AI Agents Effectively?
For Senior engineers and tech leads managing AI coding agents · Based on Priscila Andre's Comprehend-First AI Coding Skill
// TL;DR
Senior engineers and tech leads using AI coding agents face a critical risk: the agent explores the codebase, forms its own understanding, and acts on it—potentially in the wrong direction. Priscila Andre's Comprehend-First method inserts a mandatory comprehension gate between the agent's research and your planning phase. You verify what the agent found, align your mental model, and only then direct implementation. This is the difference between being an agent manager and being someone who blindly merges agent output.
Why Do AI Agents Drift During Implementation?
AI coding agents like Devin, Cursor Agent, or Claude Code can research a codebase, form a plan, and implement changes autonomously. But autonomy without oversight produces what Priscila Andre calls slop code—code shipped into production without genuine understanding of what it does or why.
The root cause of agent drift is simple: the agent's understanding of the codebase may be wrong, and if you haven't verified its research, you have no way to catch the error. You're steering blind.
Priscila Andre's framework addresses this by redefining the senior engineer's role as an agent manager—someone who primarily orchestrates and steers AI agents rather than writing code directly. The core competency is comprehension, not generation.
How Do You Insert a Comprehension Gate Into Agent Workflows?
The standard agent workflow is Research → Plan → Implement. Andre's innovation is expanding this to Research → Comprehend → Plan → Implement:
1. Research: Let the AI agent explore the codebase and gather information.
2. Comprehend: Review the agent's research output personally. Use Catch Me Up exploration modes to verify its findings. Ask: "Is this actually how the code works?" Cross-reference against your existing knowledge.
3. Plan: Only after you've verified the research, instruct the agent to plan the implementation.
4. Implement: During implementation, use your verified mental model to catch drift in real time.
The comprehension gate is the step most teams skip, and it's the step that prevents the most bugs. If you conflate research with comprehension—assuming the agent understood the code correctly because it produced a confident summary—you lose your ability to steer.
What Does Effective Agent Steering Look Like in Practice?
Consider a real scenario: your agent is tasked with refactoring a payment processing module. It researches the codebase and reports that module A calls module B directly.
Before accepting this and moving to planning, invoke Catch Me Up with Trace mode: "Trace the execution path from payment initiation to confirmation. Is there an intermediary between A and B?" You discover there's a middleware layer the agent missed. Without this verification, the agent's refactoring plan would bypass the middleware and introduce a regression.
During implementation, your verified mental model acts as your steering wheel. When the agent proposes a change that skips the middleware, you catch it immediately because you know it's wrong. This is the difference between managing an agent and being managed by one.
How Do You Track Whether You're Actually Managing Agents Well?
Priscila Andre recommends auditing your own AI usage. Categorize your last 50–100 prompts as comprehension, planning, or generation. Andre found her own ratio was 67% comprehension, 31% planning, and only 2% generation.
If your ratio is heavily skewed toward generation, you may be skipping the comprehension step. If it's skewed toward planning without comprehension, you may be planning on unverified foundations. The ideal pattern for a senior engineer managing agents is comprehension-heavy, with planning as a deliberate second step and generation as a small fraction.
What Should You Do Next?
Before your next agent-assisted task, explicitly separate research from comprehension. Let the agent research, then pause. Use Catch Me Up to verify its findings before you approve any plan. Track your prompt ratio for one week to see where you actually spend your AI interaction time. You may discover, as Andre did, that you're already doing mostly comprehension—and knowing this lets you optimize your prompts and workflows deliberately.
// FREQUENTLY ASKED QUESTIONS
How do I prevent AI agents from going in the wrong direction in my codebase?
Insert a mandatory comprehension gate between the agent's research phase and your planning phase. Review what the agent found, verify it against the actual code using Catch Me Up exploration modes, and resolve any discrepancies before approving a plan. Your verified mental model is what lets you catch agent drift during implementation.
What percentage of my AI usage should be comprehension vs. generation?
Priscila Andre found her ratio was 67% comprehension, 31% planning, and only 2% generation. While your ratio may vary, a senior engineer managing AI agents should expect comprehension to dominate. If generation is your largest category, you may be skipping the comprehension step and shipping unverified agent output.
How is being an agent manager different from traditional senior engineering?
An agent manager primarily orchestrates and steers AI agents rather than writing code directly. The core shift is from implementation skill to comprehension and quality control skill. You need deeper codebase understanding as an agent manager because you must catch errors the agent makes—which requires knowing what correct looks like before the agent starts working.