Introduction
Software development is no stranger to big shifts. Every few years something new is emerging and changing how we all build great software. AI powered coding tools aren’t just arriving—they’re already changing how experienced engineering teams work, think and deliver value. For technical leaders the question isn’t whether to adopt these tools but how to lead their teams through this change.
Recent advancements in technology, especially in deep learning and generative AI, have enabled the rise of agentic coding. These innovations allow AI tools to process unstructured data, automate complex tasks, and generate new content, transforming workflows and accelerating software development.
Agentic coding is a new paradigm where AI agents act as autonomous programming partners that can write, test and iterate on code with minimal human intervention. Agentic AI, a related concept, refers to autonomous software agents capable of reasoning, setting goals, and performing complex tasks, making them integral to process automation and collaborative workflows. Unlike simple code completion tools these agents can understand context, make architectural decisions and handle complex multi-step programming tasks. Your mileage may vary: they won’t always get it right for your business use case! Just remember that this is a rapidly evolving technology, and humans are a critical ingredient to make agentic coding work.
This guide distills insights from engineering and product teams who have gone from resistance to results and are rapidly iterating on their own agentic AI best practices. Each section links to a blog, screencast or short video with practical frameworks, real world experiences and strategic guidance for leaders ready to help their team get the productivity benefits of agentic coding.
- Code generation tools are easy to start but hard to master
- From resistance to results: Why agentic coding requires a new mindset
- The double loop model: A better approach to agentic AI
- Video Series: Uncomfortable truths about vibe coding and agentic coding
- LLMallard: A low-key AI toy app reveals the value of rubber duck sessions and thoughtful tool selection
- The business of AI: solving real problems for real people
- AI and engineering leadership: The mirror effect
- Additional Resources
Remember when this was all new? Daniel Huss shares a candid reflection on the cyclical nature of AI coding tool adoption, comparing it to the college registration experience. Explores the pattern of initial excitement, gradual disillusionment, and eventual return that many developers experience with tools like Copilot.
Key insights for leaders who have teams struggling with code generation tools
Developer relationships with AI coding tools follow predictable adoption cycles that mirror technology adoption patterns more broadly—initial excitement gives way to frustration, abandonment, and eventual re-engagement as developers learn to work within the tools' limitations. The effectiveness of these tools depends heavily on context, excelling at rote tasks and well-established patterns while struggling with esoteric problems that require domain-specific knowledge and architectural understanding.
A critical challenge emerges when AI suggestions "hijack" developer thinking, introducing what the author calls "brain worms"—plausible but suboptimal solutions that derail original problem-solving approaches and lead developers down unproductive paths. However, the tools show particular value for experienced developers working in unfamiliar languages, effectively serving as syntax translators that allow deep technical knowledge to transfer across technology stacks.
Success with AI coding tools requires developing entirely new competencies around prompt engineering—learning to constrain AI behavior through personas, context, and specific interaction patterns.
The fundamental challenge for technical leaders is ensuring their teams maintain control over these tools rather than becoming dependent on them for development decisions, which requires deliberate practice and organizational policies around appropriate usage patterns.
The universal struggle in adopting AI coding tools
Every software team faces the same fundamental questions about AI coding tools: Why do some developers get huge productivity gains and others get “junior dev AI” output? A group of experienced developers at Test Double discussed this and came to some interesting conclusions. The answer isn’t in the tools themselves, but in the mental models we bring to them. However, it's important to note that over-reliance on AI tools can diminish developers' problem solving skills, so maintaining and developing these abilities remains crucial. Participating in this discussion: Aaron Gough, Dave Mosher, Dustin Tinney, Mavrick Laakso, Rian Rainey, and Robert Komaromi.
Key takeaways for tech leaders on why AI powered tool adoption lags
The productivity paradox:
Smart developers struggle at first because they approach AI agents like human junior developers. The interaction model is fundamentally different—AI agents require specific, detailed instructions not broad problem statements.
Abstraction, not replacement:
Successful teams view agentic coding as another layer of abstraction, like how high-level languages abstracted away assembly code. The craft isn’t going away; it’s evolving to system-level thinking and architectural guidance.
The satisfaction shift:
The real satisfaction in software development comes from solved problems and positive user impact, not from writing every line of code by hand. Teams that get this unlock the full potential of agentic coding.
Business reality check:
In environments where results matter more than code purity, agentic coding gives you a competitive advantage. The key is knowing when perfect architecture matters and when working solutions delivered faster create more value.
Teams should expect resistance during AI adoption, as even senior developers will struggle if they don’t adapt their interaction patterns to work effectively with AI tools. Organizations must invest in prompt engineering capabilities, since the ability to break down problems into specific, actionable instructions becomes a core competency for success.
Productivity metrics need to be redefined when teams are delivering features in hours instead of days, requiring new approaches to sprint planning and estimation processes. Agentic coding can significantly enhance employee productivity by optimizing business processes and workflows, automating routine tasks, and enabling teams to focus on higher-value activities. Success requires teams to focus on outcomes and business results rather than engineering purity alone, shifting mindset from technical perfection to value delivery.
The problem with prescriptive AI agent prompts
Most agentic coding advice tells you to provide exhaustive context upfront—detailed PRDs, tons of documentation, and super specific requirements. However, Joé Dupuis says it's crucial to help the agent understand the problem domain and task scope, ensuring the agent comprehends both the overall objective and the specific details needed for accurate execution. Otherwise you miss AI’s greatest strength: explorability. Instead of overwhelming the agent with information, breaking down prompts into smaller, sequential steps—using previous prompt context and chained prompting—can improve focus, control, and overall agent performance. Front-loading all decisions goes against the very nature of software development.
The double loop framework for agentic coding
Loop 1: Vibe coding (Product mode)
- Iterate the feature’s look, feel, and functionality
- Don’t look at code quality yet—treat this as a spike
- Explore multiple directions in parallel, use AI’s low cost of experimentation
- Keep scope small but don’t artificially limit creativity
- Think gradient descent: iterate towards the right solution through experimentation
Loop 2: Code quality (Engineering mode)
- Once the feature feels right, open a draft PR and look at the code
- Use the AI generated code as a “McDonald’s baseline”—a starting point that triggers your expertise about what good code looks like
- Review and iterate on code quality through focused feedback
- Keep momentum by keeping scope narrow and launching multiple agents in parallel
Why the double loop works for AI agentic coding
The double loop model recognizes that software development is iterative. The planning stage is when you know the least about what you’re building. AI agents are great at rapid exploration, making spikes essentially free. This AI agentic coding approach uses that strength while maintaining code quality through experienced engineering judgment.
Use AI’s low cost iteration for exploratory programming to explore design possibilities and rapidly test different approaches. Product decisions and code quality are different problems that require different approaches, so it’s important to separate these concerns rather than conflating them.
Engineering standards must be maintained through a final loop that ensures production-ready code through experienced review and quality assurance. Teams can scale their efforts through parallelism by launching multiple agents on different problems at the same time, maximizing productivity and output.
Read the blog and watch the screencast
In this video series on the pros and cons of vibe coding, we asked our senior developers and product managers to weigh in on a variety of angles within vibe coding, agentic coding, and what this all means for building great software and great teams. Participating in the series: A.J. Hekman, Alex Martin, Dave Mosher, David Lewis, Dion Utt, Dustin Tinney, James Zimmerman, Joel Helbling, Maverick Laakso, and Tammy Lawlor.
Why vibe coding reveals more about your team than the AI tools
In this video, we shared unpopular opinions on vibe coding: provocative takes on skill requirements, ethics and developer capability vs tool effectiveness.
- Skill reflection principle: Bad results from AI coding tools are often a reflection of the developer’s lack of critical thinking and technical knowledge, not the tool’s limitations.
- Universal experimentation: All software developers should experiment with AI coding tools to know what’s possible today and track the evolution.
- Education baseline: Vibe coding requires significant foundational knowledge – it’s not a shortcut for inexperienced developers.
- Ethical considerations: Power consumption, water usage and copyright concerns are real ethical questions around widespread AI tool adoption.
- Technical debt creation: AI coding can create “clever code” problems where developers become dependent on AI for modifications to code they don’t understand.
- Dependency risk: Organizations can create situations where they can’t maintain or modify their own systems without AI.
Does AI prototyping build the right thing faster?
In this video, a panel of senior product managers and software engineers discusses whether AI accelerated prototyping leads to better discovery or premature commitment to solutions. The verdict: speed amplifies existing organizational practices—good discovery processes get faster, bad ones create more waste.
- Foundation first: AI prototyping success is entirely dependent on existing product discovery practices. Organisations without solid experimentation frameworks will build the wrong things faster.
- Collaboration risk: AI accessibility can eliminate cross-functional collaboration during prototyping, creating knowledge vacuums and risk concentration.
- Process design: Successful teams deliberately build discovery steps into their AI prototyping workflow, using design sprints before AI acceleration phases.
- Prototype vs production: Lower development costs make it easier to push prototypes to production without proper engineering rigour.
How do you balance product vibes with solving real customer needs and business problems?
Discussion on how to ensure AI-generated solutions address actual customer problems, plus the critical role of validation and the “garbage in, garbage out” principle.
- Experience dependency: Success with AI coding tools requires broad technical experience for effective pattern matching. Junior developers face significant risks without proper mentorship.
- Technical debt amplification: AI tools replicate and compound existing code quality issues—poor codebases generate worse AI suggestions.
- Validation imperative: AI-generated solutions must still undergo rigorous customer validation. Speed of creation doesn’t eliminate need for market testing.
- Data vs. intuition balance: Effective product leaders combine data-driven insights with intuitive understanding, and avoid relying solely on assumptions.
How does vibe coding (and agentic AI) impact development team productivity and code quality?
In this video, we critically examined vibe coding's effects on productivity and code quality. The conversation highlights significant risks around code review, security, and junior developer growth.
- Productivity paradox: Senior developers gain productivity while juniors may become less effective. Without proper mentoring systems, overall team productivity can decline.
- Skill atrophy risk: Core debugging and root cause analysis skills deteriorate when developers don't understand the code they're deploying.
- Security vulnerabilities: New attack vectors emerge through AI-generated dependencies and modules that may not exist or contain malicious code.
- Organizational scope: Vibe coding impacts extend beyond engineering to product and operations. Siloed implementation creates systemic problems.
- Team dynamics: Creates potential for hyper-productive developers who generate problems that less visible team members must fix, leading to burnout and retention issues.
Should you ban vibe coding anywhere in the software lifecycle?
Expert perspectives on which parts of software development should have restrictions on AI code generation, with focus on production systems, testing, and critical infrastructure.
- Production code boundaries: Strong consensus against using vibe coding for production code and tests without rigorous review processes.
- Rule of thirds: Implement pressure testing across multiple AI models and human review before accepting any AI-generated solution.
- Mission-critical question: Apply the “airplane test”—would you trust this code if your life depended on it?
- Legal and IP risks: AI-generated code creates copyright, patent, and licensing complications that can have severe business consequences.
- Maintenance reality: Someone must be able to debug and fix AI-generated code at 3 AM when external AI services are unavailable.
Is assessing quality with code smells a superpower or illusion?
Debate on whether rapid codebase quality assessment through AI is genuinely valuable or creates false confidence. Emphasizes the importance of context and experience in quality evaluation.
- Context dependency: Code quality cannot be assessed without deep organizational and historical context. Technical decisions often reflect business constraints that aren’t visible in the code.
- AI quality masking: AI-generated code can appear high-quality at surface level while hiding fundamental architectural problems.
- Consistency challenges: AI models lack consistency in their own output, potentially creating maintenance challenges across a codebase.
- Experience requirement: Rapid quality assessment is only valuable when performed by experienced developers with extensive pattern recognition abilities.
- Measurement focus: Code quality discussions should center on developer ability to reason about and extend the codebase, not subjective style preferences.
Uncomfortable truths about vibe coding
Candid discussion of uncomfortable realities around AI coding tools, including economic sustainability, job displacement, and the democratization of software creation.
- Economic sustainability: Current AI coding capabilities are heavily subsidized. Real costs (electricity, water, compute) may make human developers more economical long-term.
- Democratization impact: AI tools enable non-technical founders to create functional prototypes, potentially disrupting traditional software development business models.
- Job market evolution: The industry faces a potential transition similar to accounting after spreadsheets—not elimination but significant reduction in required workforce.
- Tool vs. solution: AI coding should be viewed as a powerful tool for existing teams rather than a replacement for engineering expertise.
- Commoditization indicator: If an entire product can be built with vibe coding, it may indicate the product lacks genuine innovation or market differentiation.
Beyond the hype: Practical AI integration with a hierarchical approach to tools
While the industry pushes toward using the most powerful (and expensive) AI models for everything, experienced teams are discovering more nuanced approaches. The key insight: different problems require different tools. Daniel Huss realized that sometimes a simple rubber duck approach beats expensive AI calls, and built a fun app to prove that out.
Local models for routine tasks:
- Simple one-liners and basic coding tasks
- Feature spikes and experimental code
- Situations where speed and cost matter more than perfection
Cloud models for complex problems:
- Large-scale refactoring
- Complex architectural decisions
- Integration with legacy systems requiring deep context
Human expertise for strategic decisions:
- System architecture and design patterns
- Business logic and domain-specific requirements
- Quality gates and production readiness
The economics of AI-assisted development
Smart teams are building sustainable practices around AI tool usage. This includes understanding when to use expensive cloud-based models versus efficient local alternatives, and recognizing that the current subsidized pricing model will inevitably shift.
Organizations should develop cost management practices that aren't dependent on artificially low pricing to ensure long-term sustainability. Teams need to focus on skill development by learning to leverage lower-fidelity models effectively, maximizing value from less expensive AI options. Tool selection becomes critical when matching capabilities to problem complexity, avoiding over-engineering solutions that don't require premium AI models. Building sustainable practices means creating workflows that remain viable as the economic landscape evolves, preparing for potential changes in AI pricing and availability.
Cathy Colliver is participating in the Perplexity AI Business Fellows program, and shares highlights from fireside talks with three business leaders who are building, implementing, or investing in AI to solve real business problems. While a lot of the AI hype cycle is missing a focus on business fundamentals, these organizations are keyed in on customer needs and taking practical approaches to meet customers where they are.
- Matt Palmer, Developer Relations at Replit talked about how they are democratizing app development for at least the throwaway prototype stage.
- Eric Glyman, Co-founder & CEO of Ramp shared insights around how they are positioning their business. Spoiler: it's about more than the tech.
- Roy Bahat, head of Bloomberg Beta discussed a useful framework for thinking about what applications of AI make sense for your business.
These three conversations illuminate a common thread: AI isn't just another technology upgrade—it’s changing how we think about capability, value creation, and competitive advantage.
AI tools don’t fix organizational problems—they amplify existing strengths and weaknesses. Teams with strong processes, clear goals, and effective leadership see dramatic productivity gains. Teams with messy workflows, technical debt, and weak fundamentals find their problems magnified. AI solutions can help organizations transform their business processes and achieve better outcomes by enabling automation, optimizing workflows, and supporting real-time decision-making.
What artificial intelligence reveals about your organization
In reality, though, AI accelerates whatever processes you have, making broken workflows painfully obvious when they’re executed at AI speed, which means process quality becomes critical. AI can also provide valuable insights into organizational strengths and weaknesses, helping leaders identify areas for improvement and optimization. Strong leadership becomes more valuable, not less, because AI requires clear direction, strategic thinking, and rigorous oversight to deliver meaningful value.
Messy codebases and poor architectural decisions create significantly more problems when AI is generating code rapidly, making addressing technical debt that much more important. Misaligned teams produce conflicting outputs faster with AI tools, while aligned teams achieve coordinated results at unprecedented speed, highlighting the crucial need for team alignment.
Key insights for technical leaders on encouraging pragmatic AI approaches
Accelerating workflows: AI handles repetitive tasks, freeing developers for high-value work. The key is having clear definitions of what constitutes high-value work.
Tightening feedback loops: AI enables faster iteration, but only when teams have strong review processes and quality gates.
Breaking down complexity: AI excels at solving well-defined problems, but requires experienced engineers to break down complex challenges appropriately.
Engineers need new skills in areas like prompt engineering, AI output validation, and system design, making it essential to invest in people who can manage AI systems in context.
Strong processes and clear goals matter more than ever, so companies should double down on establishing robust operational frameworks. Success in the AI era requires measuring business impact rather than just code output, making it crucial to focus on meaningful outcomes.
Conclusion: Leading the agentic coding future
Agentic coding is already impacting the pace and quality of building new software products and refactoring existing legacy codebases.
Engineering leaders should understand that AI is not replacing human expertise. In reality, the likely path forward is to amplify AI tools with strategic and pragmatic guidance by experienced developers and product managers.
Success requires more than just adopting new AI tools. It demands a fundamental shift in how we think about software development, team leadership, and value creation. The most successful organizations will be those that:
- Embrace the learning curve: Even experienced developers need time to develop new mental models for working with AI—support that with growth time
- Invest in fundamentals: Strong processes, clear communication, and solid technical foundations become more important, not less
- Focus on outcomes: Measure success by business impact and user value delivered
- Build sustainable practices: Develop flexible approaches that remain viable as agentic coding and other AI tools evolve
Judgment, creativity, and strategic thinking make great software possible, and can help your team get the most out of AI tools. Software developers and product managers who have been-there-done-that across a wide-range of challenges, situations, and domains are even more valuable in our new reality of agentic coding.
Test Double Pragmatic AI YouTube Playlist
Pragmatic AI Videos from Test Double: Playlist of videos exploring practical implications of vibe coding, plus a screencast showcasing advanced techniques in agentic coding with The Double Loop Model.
Resources
Below are resources our team of senior software developers and senior product managers are exploring in the moment we are in, July 2025. This is by no means an exhaustive list of tools and resources, and we guarantee it's going to shift and evolve as we continue experimenting and learning.
Editors and tools
- Claude Code
- Opencode
- Junie
- Linear
- Void
- Zed
- Open VSX Registry
- Tonkotsu
- Magistral
- Jules
- superwhisper
- LM Studio
- Drata AI-Native Platform
- Warp
- gemini-cli
- mistral-code
Model features
- Claude 4 breakdown from AI Explained
- Artificial Analysis’s comparison of models: Intelligence, performance & price analysis
- Welch Labs How DeepSeek rewrote the transformer
Local LLM
MCPs
MCP servers can be installed and configured to connect, configure, and secure AI agent tools, facilitating integration with internal data sources like Confluence, Jira, Notion, or GitHub. The mcp server enables agents to access organizational resources, coding standards, and documentation by connecting these repositories through configuration files, supporting secure and platform-approved development workflows.
- Glama MCP hosting platform
- Language service index format overview
- Interfacting MCP with Combinatorial, Convex and SMT Solvers
- Joel Helbling: MCpeasy
Design resources
- Foundation Capital & Designer Fund: State of AI in design
- ESLint dashboard preview | Claude
- AI Engineer: The RAG stack we landed on after 37 fails
- I map out every single file before coding and it changed everything
Learning resources
- How Anthropic teams use Claude Code
- Mastering Claude Code in 30 minutes
- Accelerating large-scale test migration with LLMs
- Nextra: Prompt engineering guide
- The remote AI runtime for Phoenix
- Gandalf
- How AI models learn, explained visually
This guide synthesizes insights from the experienced software engineering and product management teams at Test Double and industry research. For the latest updates and additional resources, check out the Test Double insights blog or subscribe to the Test Double Dispatch newsletter.