Agentic Coding

Briefing: Generative AI Insights Purpose: I'm a software engineer who's looking to stay up to date with developments in the generative AI (gen AI) space. As an early-stage startup developer, my primary focus is building on top of an LLM based system. Some topics I'm interested include: - Context Engineering techniques; especially those that mitigate hallucinations, summaries, and accurate quotations. - Best practices when it comes to coding agents - Macro information about the business of running an early stage startup focusing on gen AI products - Using LLM models to translate to different languages with high accuracy and correctness - Application memory (short-term, long-term, retrieval, user profiles) in real products. - LLM evals and monitoring: automated tests, metrics, and product-level evaluation loops.

Key Insights

• Context engineering is evolving into a distinct discipline with multiple specialized layers. Moving beyond simple prompt crafting, "context engineering" now encompasses curating the entire information environment an agent operates within, including system prompts, tool definitions, and retrieved documents. This is further segmented into higher-level skills like "intent engineering" (encoding organizational goals) and "specification engineering" (creating documents agents can execute against autonomously), which are becoming critical as models gain capability and operate over longer time horizons.
  • 'Prompting' Just Split Into 4 Skills. You Only Know One. Here's Why You Need the Other 3 in 2026.
  • The Powerful Alternative To Fine-Tuning
• A "human-in-the-loop" approach is essential for mitigating hallucinations and ensuring accuracy in complex agent workflows. While automated systems are improving, human oversight remains critical, particularly for tasks requiring precise specification or judgment. Techniques such as "vibe coding," where humans iteratively refine specifications based on AI output, and "failure model maintenance," which involves understanding specific agent weaknesses, are key to leveraging AI effectively without succumbing to errors or "context rot."
  • My 10-Year-Old Vibe Codes. She Also Does Math by Hand. Why That's the Only Strategy That Works.
  • Why Every AI Skill You Learned 6 Months Ago Is Already Wrong (And What Is Replacing Them)
  • The Case for Letting Your AI Forget
• The business landscape for gen AI startups is shifting towards "agent-native" growth and lean operations. Startups are finding success by focusing on deep user engagement and retention ("relationship effects") rather than viral growth, which can starve the feedback loops necessary for product improvement. Additionally, the rise of "tiny teams" or even solo founders leveraging AI agents to automate tasks like coding, marketing, and customer support suggests a new model for achieving significant scale with minimal headcount.
  • You Should Never Go Viral With Your AI App
  • ⚡️ Polsia: Solo Founder Tiny Team from 0 to 1m ARR in 1 month & the future of Self-Running Companies
  • Hot take: solo founders with AI are about to build stuff faster than small teams
• Automated evaluation and "seam design" are becoming critical for reliable AI product development. As AI capabilities advance rapidly, traditional metrics and manual testing are insufficient. Concepts like "seam design" (structuring work for clean handoffs between humans and AI) and rigorous "evaluation design" (creating measurable proof of quality) are emerging as essential practices. Furthermore, tools that automate context optimization and generate reasoning strategies are providing a cost-effective alternative to training new models from scratch.
  • Why Every AI Skill You Learned 6 Months Ago Is Already Wrong (And What Is Replacing Them)
  • 'Prompting' Just Split Into 4 Skills. You Only Know One. Here's Why You Need the Other 3 in 2026.
  • The Powerful Alternative To Fine-Tuning
• Security and permission models for AI agents are critical, unresolved infrastructure challenges. Current permission models for coding agents are often binary (all-or-nothing), leading to security risks or usability issues. A robust future requires granular, declarative policies and relationship-based scoping, moving security from simple dialog boxes to comprehensive infrastructure solutions like OpenFGA.
  • Claude Code's broken permission model

Emerging Patterns

• The shift from prompt engineering to "specification engineering" and "context curation." Multiple sources highlight a move away from tweaking individual prompts towards engineering the broader context and specifications that guide agent behavior. This involves creating structured documents (like CLAUDE.md or SKILL.md) and managing the entire information environment to ensure consistency and reliability, especially for complex, long-running tasks.
  • 'Prompting' Just Split Into 4 Skills. You Only Know One. Here's Why You Need the Other 3 in 2026.
  • stopped fighting Claude Code after I actually wrote a proper CLAUDE.md
  • Video-to-skill pipeline: turning YouTube tutorials into Claude Code context with OCR + two-pass AI enhancement
• The debate between CLI tools vs. specialized protocols like MCP for agent integration. While some advocate for specialized protocols like the Model Context Protocol (MCP) to standardize tool use, others argue that simple, composable CLI tools are often more practical, debuggable, and reliable for agentic workflows. This tension reflects a broader discussion on the most effective way to equip AI agents with capabilities.
  • When does MCP make sense vs CLI?
  • 'Prompting' Just Split Into 4 Skills. You Only Know One. Here's Why You Need the Other 3 in 2026.

Dissenting Views

• Skepticism regarding the utility of massive context windows. While some celebrate the arrival of 1 million token context windows, others question their practical value, suggesting that performance degrades and costs rise significantly with such large contexts. This view contradicts the general excitement around larger context windows, arguing instead for more targeted context management or the use of sub-agents.
  • Got the 1 Mil Context Window. 5x Plan. Did ya'll get it? 🤩
• The effectiveness of AI-driven "solo founders" is debated. While some see AI enabling solo founders to compete with larger teams, others argue that this merely raises the bar for MVPs and that the true value of a startup lies in areas that AI cannot easily replicate or scale, such as team collaboration and execution beyond just code generation.
  • Hot take: solo founders with AI are about to build stuff faster than small teams

Read & Act

What to read

• 'Prompting' Just Split Into 4 Skills. You Only Know One. Here's Why You Need the Other 3 in 2026. — This video provides a crucial framework for understanding the evolution of prompting into distinct engineering disciplines (context, intent, specification). It's essential for moving beyond basic interactions to building robust, agent-based systems.
• Claude Code's broken permission model — A critical read for any developer building coding agents, as it highlights significant security flaws in current permission models and proposes a necessary shift towards infrastructure-level authorization.
• The Powerful Alternative To Fine-Tuning — This offers a compelling alternative strategy for improving model performance without the high costs of fine-tuning, relevant for startups looking to optimize their AI systems economically.
• You Should Never Go Viral With Your AI App — This article provides a counter-intuitive but strategic perspective on growing AI products, emphasizing retention and relationship building over rapid user acquisition, which is vital for long-term viability.

What to do

• Implement a structured context file (e.g., CLAUDE.md) for your coding projects. Define clear rules, coding conventions, and constraints to guide your AI agent's behavior. Experiment with placing these files in specific subdirectories to tailor agent responses for different parts of your codebase (frontend vs. backend).
• Develop a "failure model maintenance" routine. Instead of just fixing bugs as they arise, actively document and categorize the specific ways your AI agents fail. Use this to refine your "seam design"—the points where human oversight or specific verification steps are inserted into the workflow to catch these errors.
• Experiment with "specification engineering" for complex tasks. Before coding, practice writing self-contained problem statements and clear acceptance criteria that an agent can execute against. This shifts your focus from improved prompting to improved task definition, a higher-leverage activity for agent-based development.