Agentic Engineering: Working With AI, Not Just Using It — Brendan O'Leary

Key Concepts

• Agentic Engineering: A paradigm shift in software development where AI is treated as a collaborative partner rather than a mere tool (like autocomplete).
• Context Engineering: The deliberate management of information provided to an AI agent to ensure high-quality outputs while minimizing costs and "context poisoning."
• Research-Plan-Implement Loop: A structured methodology to prevent AI from generating "garbage code" by prioritizing problem understanding and planning before execution.
• Model Context Protocol (MCP): A standard for connecting AI agents to external data sources, APIs, and tools.
• Mental Model: Viewing the AI agent as an "enthusiastic, well-read, but judgment-lacking junior developer."

1. The Evolution of AI in Engineering

The speaker outlines the rapid progression of AI tools:

• Early 2020s: "Autocomplete on steroids"—AI suggested single lines of code.
• 2022: Function-level suggestions (e.g., GitHub Copilot), moving AI from novelty to utility.
• 2025–2026: The era of Agentic Engineering, where models can execute tasks, modify multiple files, run tests, and submit pull requests.
• Core Argument: We are no longer just "using" machines; we are "working with" them. Success depends on the engineer’s ability to direct the AI, not just accept its output.

2. The "Junior Developer" Mental Model

The speaker argues that AI agents should be treated like a highly capable but inexperienced junior developer:

• Strengths: Incredibly fast, tireless, no ego, and possesses vast knowledge of frameworks and patterns.
• Weaknesses: Lacks business context, lacks architectural judgment, and can be "confidently wrong."
• Actionable Insight: An engineer’s value lies in providing the "judgment" and "context" that the agent lacks.

3. Context Engineering: The Art and Science

Context is both expensive (token costs) and volatile (quality degradation).

• The "Dumb Zone": Filling a context window beyond 50% often leads to a decline in model reasoning quality.
• Context Poisoning: Occurs when outdated comments, irrelevant MCP servers, or failed previous attempts remain in the session, causing the agent to repeat mistakes.
• Best Practices:
  • Persist outside the window: Use agents.md or memory files to store project-wide rules.
  • Isolate tasks: Use parallel agents or separate sessions for distinct tasks to prevent context accumulation.
  • Summarize and Trim: When a session becomes complex, ask the agent to summarize the current state, then start a fresh session with that summary.

4. Methodology: The Research-Plan-Implement Loop

To avoid the "garbage in, garbage out" trap, the speaker proposes a three-phase workflow:

1. Research (Ask Mode): Use a restricted mode where the agent cannot write files. Focus on understanding the codebase, data flow, and edge cases.
2. Plan: Create a step-by-step plan.md file. This includes specific file changes, test verification strategies, and clear definitions of scope.
3. Implement: Execute the plan in a new, clean session. This allows for smaller, cheaper models to perform the work because the "hard thinking" is already done.

5. Configuration and Tooling

• agents.md vs. skills.md:
  • agents.md: Always-on project rules, conventions, and testing requirements.
  • skills.md: On-demand, reusable playbooks for specific workflows (e.g., compiling change logs).
• MCP Servers: Use them to connect to GitHub APIs or documentation, but disable unused servers to save tokens and prevent the agent from getting "confused" by irrelevant tools.
• Git as a Review Tool: Treat local Git commits as a "first-pass code review" before submitting a formal pull request.

6. Notable Quotes

• "We're no longer just using machines. We're now working with them." — Attributed to Armin (creator of Flask).
• "AI can't replace thinking. It can only amplify the thinking you've done or the lack of thinking you haven't done." — Attributed to Dex Horthy.
• "A bad line of research can potentially be hundreds of lines of bad code." — Attributed to Dex Horthy.

Synthesis and Conclusion

The transition to agentic engineering requires a shift from "coding" to "directing." By adopting a structured Research-Plan-Implement framework and treating the AI as a junior developer requiring clear, curated context, engineers can reclaim significant time (up to 30% as noted by the speaker). The ultimate goal is to offload tedious tasks to the agent, allowing human engineers to focus on high-level architectural decisions and complex problem-solving.