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, Best practices for coding agents, Macro business info for startups, Application memory, and LLM evals.

Key Insights

• Agent error modes are shifting from hallucinations to omissions. Data from 5,000+ quality checks on autonomous coding agents reveals that "forgetting" tasks or leaving stubs (e.g., // TODO) is now a more prevalent issue than hallucination. This is attributed to "context compression" where nested dependencies are dropped as the context window fills; the mitigation is breaking work into bounded tasks with fresh contexts rather than one long thread.

• 97 days running autonomous Claude Code agents with 5,109 quality checks. Here's what actually breaks.

• Engineering is no longer the primary bottleneck, changing startup risk profiles. As AI reduces the cost of coding and shipping features, the historic business justification for expensive risk mitigation (extensive pre-release testing) weakens. Startups may shift optimization targets from "stability" to "decision speed," consolidating decision power into fewer individuals driving agentic machines to maximize throughput.

• AI Moves the Bottleneck - Are You Ready for What That Means For Your Career?

• Workflow is evolving from "Prompting" to "Blueprinting." Effective agent orchestration now requires defining visual "blueprints" or workflows—anticipating decision branches and data gaps—before engaging the model. This separates the planning phase (architecture/workflow) from execution, allowing agents to run complex background tasks without constant human-in-the-loop micromanagement.

• Not Prompts, Blueprints

• Touched Grass: 0/73 days. How I Use Claude Code

• Evaluation is moving to the "Skill" level, not just the Model level. New tooling (like Anthropic’s skill-creator) now supports defining automated evaluations for specific tool-use definitions ("skills"). This marks a maturity shift where context and skills are treated as software requiring unit tests, moving away from subjective "vibes-based" assessment of whether a prompt "feels helpful."

• Claude brings evaluations to their skills

• Startup advantage varies drastically by stage. Seed-stage founders report engineering velocity increases of up to 10x using AI, whereas Series D companies report only marginal gains (~10%). This discrepancy suggests early-stage startups have a distinct, perhaps temporary, asymmetric advantage in product velocity compared to incumbents.

• How Claws Took Over Every

Emerging Patterns

• Cognitive Load Transfer: From Syntax to Orchestration. While AI accelerates code generation, developers report a spike in mental exhaustion and burnout. The cognitive load has shifted from implementation details to managing multiple parallel "agent" threads and reviewing complex outputs. This necessitates new personal workflows, such as strict session management or treating agents like junior employees that require "progressive trust."
• AI coding helps me with speed, but the mental overload is heavy! How do you deal with it?
• 🎙️ This week on How I AI: 5 OpenClaw agents run my home, finances, and code & How Coinbase scaled AI to 1,000+ en…

• The last months be like

• Legal Risks in AI-Assisted Refactoring. A growing concern involves "license laundering," where developers use AI to rewrite Copyleft (e.g., LGPL) code into permissive (e.g., MIT) formats. This raises unresolved legal questions about derivative works, posing a hidden liability for startups incorporating AI-refactored open-source components into their proprietary stacks.

• Relicensing with AI-Assisted Rewrite

Read & Act

What to read

• 97 days running autonomous Claude Code agents with 5,109 quality checks. Here's what actually breaks. — Essential reading for your interest in context engineering. It provides empirical data on why agents fail (context compression) and offers concrete architectural patterns to fix it.
• AI Moves the Bottleneck - Are You Ready for What That Means For Your Career? — A strong macro-business analysis. It frames the strategic shift you need to make as an early-stage founder: optimizing for decision speed rather than just engineering throughput.
• Claude brings evaluations to their skills — Highly relevant for your interest in LLM evals. It discusses the specific implementation of testing "skills" (tool definitions) to prevent regression in agent capabilities.

What to do

• Implement "Context Partitioning" in your agent architecture. Instead of feeding an agent the entire project history, break workflows into bounded tasks. Initiate fresh context windows for distinct sub-tasks to prevent the "omission error" pattern caused by context compression.
• Formalize "Blueprints" for complex tasks. Before coding, sketch the workflow logic (decision trees, data requirements) visually or in pseudocode. Use this artifact to prime the agent, rather than relying on iterative conversational prompting.
• Establish an "Eval-First" workflow for tool definitions. If you are building custom tools/skills for your LLM, define the success metrics and test cases before implementation. Avoid the "feels helpful" trap by creating reproducible test scenarios (e.g., CLAUDE.local.md configurations) to verify tool performance.