How to Get Better Results From AI Coding Agents

TL;DR

AI coding agents can produce high-quality code when experienced engineers manage context carefully, guide workflows, and evaluate the results. For large brownfield codebases, consistent team practices matter as much as individual output quality, while deterministic workflow steps can make research, planning, and implementation more reliable and easier to adopt across an organization.

Transcript

Amazing. >> I think we're Yeah, we're definitely on. >> What's up, folks? I'm here with Dex. You prefer Dex or Dexter? >> I go with either. I'm very, you know, they say good engineers are lazy. Dex is half the syllables, so I go with Dex off often. >> Very good. Yeah. Jit the T, I suppose. Um, >> yeah. >> Uh, I'm Matt PCO. Um, we're going to talk a... Read More

Key Insights

• AI can produce very good code when engineers know how to guide the system and assess its output. People without programming experience may find it difficult both to request strong implementations and to recognize whether the resulting code is sound.
• AI-assisted engineering still requires thoughtful system design. Having an API generate code does not remove the need to reason about architecture, collaboration, constraints, or the difficult problems that determine whether software works effectively.
• The best reported results for large brownfield codebases are roughly two to three times faster on many tasks. Dex frames this as a current best-in-class outcome for experienced engineers working in established repositories rather than as automatic productivity for every user.
• Organizational consistency is essential even when individual engineers generate excellent code. If one hundred capable engineers use entirely different AI workflows, the resulting codebase and development process can become chaotic despite the quality of each isolated contribution.
• Context engineering is more important than finding one perfect prompt. Dex emphasizes intentional context-window management and compaction because agent performance changes with the quantity and quality of information placed into the model's working context.
• Research, plan, and implement is not a permanent formula. Dex originally expected its wording and number of stages to evolve, and his team later concluded that practical workflows needed six steps instead of the earlier three-step structure.
• Deterministic control flow makes known workflows more reliable. When tasks must happen in a specific order, software can enforce that sequence instead of embedding approximately fifty instructions inside a single planning prompt and hoping the agent follows them consistently.
• Guided workflows reduce the amount of prompt expertise required from users. Splitting a long planning process into smaller stages lets engineers participate in consequential decisions without needing to remember special wording at precise moments in the interaction.

Questions & Answers

Q: Can AI coding agents produce high-quality code?

AI coding agents can produce very good code, but the engineer must know how to guide them and evaluate what they generate. Programming experience helps users recognize weak implementations, supply relevant context, and make sound design decisions. Code generation does not eliminate engineering work, because teams still need to design systems, solve difficult problems, and ensure that generated changes fit the existing codebase.

Q: How should experienced engineers use AI coding agents?

Experienced engineers should treat AI coding agents as tools that require deliberate context management, structured workflows, and informed review. The engineer remains responsible for architecture, system design, and difficult decisions. Dex's stated goal is to help staff and principal engineers perform their work faster while reserving their attention for high-leverage activities such as shipping, designing systems, and solving hard engineering problems.

Q: Why is AI coding difficult in large brownfield codebases?

Large brownfield codebases contain established systems and extensive existing context that an agent must understand before making useful changes. Dex specifically discusses companies with millions of lines of code and thousands of engineers. In that environment, success requires more than generating plausible code. The work must fit existing architecture, follow organizational practices, and remain consistent with changes produced by many other engineers.

Q: How much faster can AI make brownfield software development?

Dex describes two to three times faster performance on most tasks as a best-in-class result for brownfield codebases at the time of the conversation. This figure is presented as an outcome for capable engineers using AI effectively, not as a guaranteed improvement for every person or task. Achieving it still requires careful context management, sound engineering judgment, and workflows suited to established repositories.

Q: Why do engineering teams need standardized AI workflows?

Engineering teams need standardized AI workflows because excellent individual outputs can still produce organizational chaos when everyone works differently. Dex gives the example of one hundred engineers who all generate code matching what they would have written manually, but who use incompatible methods. Shared practices, platforms, and workflow standards help teams coordinate changes and make AI-assisted development repeatable across the organization.

Q: What is context engineering for AI coding agents?

Context engineering is the deliberate management of the information available inside an AI agent's context window. Dex argues that the central challenge is not discovering a perfect prompt, but deciding what context to provide and when to compact it intentionally. The conversation also introduces smart and dumb zones, reflecting the idea that agent performance can vary as the context window fills with information.

Q: Why did the research-plan-implement workflow expand?

The research, plan, and implement workflow expanded because real customer use showed that three broad stages were not sufficient for consistently strong results. Dex's team observed that people could practice the prompts extensively and achieve excellent outcomes, yet teammates given the same prompts produced widely varying quality. The redesigned workflow therefore uses six steps and provides more guidance throughout the process.

Q: Why use deterministic control flow instead of one long prompt?

Deterministic control flow is preferable when the sequence of a workflow is already known because software can guarantee that steps occur in the intended order. Dex notes that the open-source create-plan prompt contains about fifty instructions. Splitting that planning process into multiple controlled stages increases reliability and reduces dependence on users remembering special phrases or mastering several complicated prompts.

Summary & Key Takeaways

• AI-generated code is not inherently poor quality. Strong results depend on the engineer's ability to guide the agent and judge its work. Dex focuses on helping staff and principal engineers use AI effectively within large brownfield codebases, where established systems, organizational constraints, and millions of lines of code complicate development.

• Individual proficiency is insufficient when engineers across an organization use incompatible methods. Even if every person produces excellent code, inconsistent workflows can create chaos. Teams therefore need shared practices, platforms, and standards that preserve code quality while making AI-assisted development repeatable across hundreds or thousands of engineers.

• Research, plan, and implement was a useful initial workflow, but practical adoption revealed the need for additional guided steps. Rather than expecting users to master several lengthy prompts, Dex's team is moving instructions into deterministic control flow, helping users focus their attention on consequential design decisions and other high-leverage engineering work.