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 when it comes to coding agents, Macro information about the business of running an early stage startup, and LLM evals/monitoring.

Key Insights

• Agentic workflows are shifting from static chains to dynamic "rivalry" and runtime coordination. Developers are moving beyond pre-defined chains (like LangChain) toward architectures where agents dynamically determine their own topology. Techniques include "agentic braintrusts" where one model drafts a plan and rival models (e.g., Gemini vs. Claude) critique it before implementation. Furthermore, new frameworks allow agents to choose at runtime whether to spawn (start fresh) or fork (inherit context) sub-agents based on the complexity of the task.
• 9 Observations from Building with AI Agents

• Cord: Coordinating Trees of AI Agents

• Context engineering is evolving into "Context Curation" and tiered memory architectures. To mitigate hallucinations and context window overflow, engineers are using clustering algorithms and decision trees to identify high-value context rather than dumping raw data. A clearer memory architecture is emerging for personalization: separating "Defaults" files (opinionated baselines for quality) from "Taste" files (user-specific preferences that accumulate over time), preventing user data from corrupting core system instructions.

• How Luxury Handbags Can Help Solve AI's Context Problem

• What Board Games Taught Me About Working with AI

• Self-correction and "Spec-First" generation are becoming standard for code accuracy. Reliable code generation now relies on a "Plan, Build, Verify" loop rather than zero-shot prompting. New tools generate detailed specification documents (e.g., SPEC.md) or visual schemas first, which agents then use as a ground-truth contract to verify their own implementation. Additionally, using languages with static typing (like Rust) is being leveraged as a "compiler-level spell check" to mechanically prevent hallucinated syntax.

• Claude Code's Superpowers plugin actually delivers
• I'm making a visual tool to create detailed specs and designs before handing off to Claude for building. (There's free credits, can also give more free access - DM me!)

• 9 Observations from Building with AI Agents

• Domain-specific ontologies are proving more effective than generic reasoning for technical accuracy. For applications requiring high precision (like translation or technical transcription), relying solely on LLM reasoning often fails. A successful pattern involves preprocessing input against strict "engineering ontologies" or dictionaries to validate terms before the model generates final output. This "dictionary-first" pipeline significantly reduces hallucinations in jargon-heavy environments.

• Show HN: Script Snap – Extract code from videos

• The business of GenAI is pivoting to trust-based monetization and "thicker" work. As models become commodities competing on cost (indicated by benchmarks like Tau2), startups are shifting business models from ad-supported to subscription/enterprise to maintain user trust. Simultaneously, the "efficiency" narrative is being challenged; rather than reducing hours, AI is creating "thicker" deliverables (e.g., 5x longer contracts because both sides use AI to analyze every clause), changing the value prop from "time saved" to "output quality/density."

• 9 Observations from Building with AI Agents
• [Scout] Trust and Visibility Shifts Reshape AI Feeds
• Does AI really save time?

Emerging Patterns

The "LLM-as-Judge" Evaluation Loop Multiple sources point to the automation of evaluation and optimization. Systems are now being built to capture agent failures (timeouts, corrections), which are then fed into a separate "LLM-as-judge" to rewrite the system prompts automatically. This creates a closed-loop system that improves weekly without manual engineering intervention. - 9 Observations from Building with AI Agents - This OpenClaw Agent Replaced 45 Minutes of Video Editing

The Rise of "Visual-to-Code" Intermediaries There is a distinct trend toward using visual interfaces not just for design, but as structured input generators for agents. Whether it's a "canvas" that outputs a Markdown spec or block-based logic for kids that triggers backend coding agents, the pattern is to abstract the prompt engineering behind a visual UI to ensure better context for the LLM. - I'm making a visual tool to create detailed specs and designs before handing off to Claude for building. (There's free credits, can also give more free access - DM me!) - Built with Opus 4.6 a Claude Code Hackathon Winners Announced

Dissenting Views

"Coding is Solved" vs. "Adoption is Slow" Boris Cherny (Head of Claude Code) posits that coding is effectively "solved" and predicts a massive shift where engineers stop writing code manually, supported by data showing rapid adoption of agentic tools. Conversely, other industry observers argue this view is hyped, noting that organizational speed, complex "human" processes (like bureaucracy), and the "S-curve" of technology adoption mean the actual economic impact and displacement of work will be much slower and less absolute than predicted. - Head of Claude Code: What happens after coding is solved | Boris Cherny - Does AI really save time?

Read & Act

What to read

• Cord: Coordinating Trees of AI Agents — This is a critical read for understanding the next generation of agent architecture. It challenges the standard "pre-defined graph" approach (like LangGraph) and demonstrates how to let agents dynamically build their own execution trees at runtime.
• 9 Observations from Building with AI Agents — A high-density resource for engineering best practices. It covers practical tactics like "rival agents," using static typing to catch hallucinations, and setting up automated prompt optimization loops.
• How Luxury Handbags Can Help Solve AI's Context Problem — Offers a clear methodology for "Context Engineering." It moves beyond abstract advice and explains how to use clustering and decision trees to curate the exact context needed for an agent to perform well.

What to do

• Implement a "Plan vs. Implementation" Check: If you are building coding agents, do not rely on a single pass. Architect a workflow where Agent A writes a spec/plan, Agent B writes the code, and Agent C (or a script) verifies the code specifically against Agent A's plan.
• Separate "Defaults" from "Taste": In your application memory, architect user profiles to store preferences (Taste) separately from system instructions (Defaults). This allows you to update core system behavior without breaking user personalization, and prevents user inputs from overriding safety guardrails.
• Build an "LLM-as-Judge" Loop: Stop manually tweaking prompts based on anecdotal failures. Log your agent's failed outputs (timeouts, user corrections) and build a nightly job where an LLM analyzes these logs to suggest prompt improvements automatically.