How Do Senior Engineers Steer AI Agents With Comprehend-First?
For Senior engineers and tech leads managing AI agents in large codebases · Based on Priscila Andre's Comprehend-First AI Coding Skill
// TL;DR
For senior engineers managing AI agents in large codebases, Priscila Andre's Comprehend-First framework is a steering methodology. The four-phase model — Research → Comprehend → Plan → Implement — ensures you personally understand the agent's research output before authorizing planning or implementation. Without this comprehension gate, agents drift, produce slop code, and make incorrect architectural assumptions. The framework positions you as an 'agent manager' whose primary value is comprehension, direction-setting, and quality control rather than raw code production.
Why do AI agents produce slop code in large codebases?
AI agents are powerful researchers but unreliable decision-makers in complex systems. When given a task in a large codebase, an agent will scan files, identify patterns, and propose solutions — but it does this without the institutional context, architectural intent, or historical reasoning that a senior engineer carries. If you skip the comprehension step and let the agent plan and implement autonomously, it will often go in the wrong direction confidently.
Priscila Andre calls this the fundamental risk of agentic coding: the agent can do the research, but you must comprehend its findings before it plans. The commonly cited three-phase model of Research → Plan → Implement is missing the critical comprehension gate.
How do I insert a comprehension gate into agentic workflows?
After the agent completes its research phase, stop the workflow. Use the Catch Me Up prompt to structure the agent's findings into something you can verify:
1. Declare your context: "I'm the tech lead for this subsystem. The agent just researched how to implement feature X."
2. Select exploration modes: Use Trace to follow the data path the agent plans to modify. Use Architecture to verify the agent's understanding of component relationships. Use Convention to check if the proposed approach follows existing patterns.
3. Demand structured output: Ask for a table comparing the agent's proposed approach against existing patterns. Request a before/after flow diagram.
4. Interrogate discrepancies: If the agent's model contradicts your understanding, this is where bugs and misalignments surface. Resolve them before authorizing planning.
Only after you can explain the approach in plain language — and would be comfortable presenting it at a keynote — do you move the agent to planning and implementation.
What does being an 'agent manager' actually look like day-to-day?
Priscila Andre describes the shift: your primary output is no longer code — it's comprehension, direction, and quality control. In practice, this means:
- 67% of your AI prompts are comprehension. You are constantly building and rebuilding mental models as the codebase evolves.
- Planning prompts come second. You translate your comprehension into specific, well-scoped implementation instructions for the agent.
- Only 2% is direct code generation. The agent handles implementation, but you steer it continuously based on your mental model.
Track your own AI usage patterns to see if this ratio matches your workflow. Most engineers are surprised to discover how little of their AI usage is actually generation.
How do I prevent agent drift during implementation?
Your comprehension from the Catch Me Up phase is your steering wheel. During implementation, continuously compare the agent's output against your mental model. When the agent proposes something that contradicts your understanding — a wrong import path, an incorrect assumption about data flow, a violation of naming conventions — surface the conflict immediately.
The Comprehend-First framework makes this possible because you have a verified model to compare against. Without it, you're rubber-stamping agent output, which is how slop code enters production.
Next step: Audit your last 20 AI prompts. Categorize each as comprehension, planning, generation, or debugging. Use the results to calibrate how much you're investing in comprehension versus generation — and whether the ratio is serving your codebase well.
// FREQUENTLY ASKED QUESTIONS
How do I know when an AI agent has done enough research before I review it?
The agent has done enough research when it can answer your specific comprehension questions about the subsystem it's about to modify. Invoke Catch Me Up against the agent's research output. If the answers are consistent and match your partial knowledge, proceed to planning. If you find gaps or contradictions, send the agent back for more targeted research.
Can Comprehend-First work with autonomous coding agents like Devin or Cursor Agent?
Yes, but you must insert manual checkpoints. Configure the agent to pause after its research phase and before planning. Use the Catch Me Up prompt to verify the agent's understanding at each checkpoint. Fully autonomous pipelines without comprehension gates are the primary source of slop code in agentic workflows.
What percentage of my AI usage should be comprehension versus generation?
Priscila Andre found her split was 67% comprehension and 2% generation. Your ratio will vary, but if generation dominates your usage in a large codebase, you may be skipping the comprehension step. Track your prompts for a week and categorize them. Most senior engineers find comprehension is — and should be — the majority of their AI interaction.