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, Coding Agents, Startup Business Strategy, Translation, Application Memory, and Evals.

Key Insights

• The "Per-Seat" SaaS Model is Collapsing into "Agentic Architecture." The industry is witnessing a shift away from pricing models based on human logins, as AI agents increasingly perform the work. Startups are advised to pivot toward charging for data value, accountability, and outcomes rather than headcount. This "resource allocation crisis" means engineers at incumbent firms are stuck maintaining legacy per-seat systems, creating an opening for AI-native startups to build "agent-first" architectures from the ground up.

  • The $285 Billion Crash Wall Street Won't Explain Honestly. Here's What Everyone Missed.
  • The Week in Charts (2/11/26)

• Agent Memory is Evolving Beyond Vector Databases. For coding agents specifically, complex vector stores and knowledge graphs may be "unnecessary overhead." A proposed five-layer memory taxonomy (Working, Core, Procedural, Archival, Recall) suggests that simple SQLite databases with Full-Text Search (FTS5) often outperform embedding pipelines for "Recall Memory." This is because LLMs can reason to construct targeted keyword queries, making semantic search less critical for code retrieval.

  • What I Learned Building a Memory System for My Coding Agent

• "Compound Engineering" and Structured Context Reduce Token Waste. To manage costs and improve accuracy, developers are moving away from dumping raw files into context. Techniques like "context-packs"—structured summaries of repo maps, key decisions, and APIs—and Claude.md files allow agents to retain "Core Memory" without re-reading the entire codebase. This aligns with the "Compound Engineering" methodology, where 50% of time is spent on the "meta-layer" (planning and system design) to simplify the agent's execution tasks.

  • How are you managing multiple Claude sessions without hitting limits?
  • Add docs and comment your code - done
  • Compound Engineering Explained

• Multi-Agent "Debate" Loops Improve Code Quality. Single-model generation is being superseded by multi-agent workflows where one model generates and another reviews or "debates" the architecture. Tools like "CodeMoot" demonstrate that having different models (e.g., Claude and Codex) critique each other before final implementation significantly reduces bugs. This mimics human code review processes but occurs autonomously, allowing for "software-shaped intent" where the human directs the outcome rather than the syntax.

  • I automated the Claude Code and codex workflow into a single CLI tool: they debate, review, and fix code together
  • Going Slower Feels Safer, But Your Domain Expertise Won't Save You Anymore. Here's What Will.
  • Claude Opus 4.6: The Biggest AI Jump I've Covered--It's Not Close. (Here's What You Need to Know)

• Evaluation Frameworks are Moving to "Minimum Viable Quality" (MVQ). To mitigate hallucinations in product environments, developers are defining MVQ thresholds: an "acceptable bar," a "delight bar," and a "do-not-ship bar." A practical technique for setting these guardrails involves prompting the model for an "ideal" response to chaotic input and comparing it against the actual output to identify specific failure modes. This helps determine where explicit constraints or "software-shaped intent" must be injected to prevent the model from inventing structure where none exists.

  • Building AI product sense, part 2

Emerging Patterns

• Specialized Pre-Training over Fine-Tuning. There is a move toward embedding core capabilities (like reasoning, self-correction, and safety) directly into the foundation model's pre-training or mid-training phases, rather than relying on post-training fine-tuning. Evidence includes models like "Almo 3.1" and newer GPT series displaying "internal turmoil" or self-correction behaviors (e.g., the "seahorse emoji" dilemma) that emerge from training data containing reasoning traces.

  • ⚡️ Reverse Engineering OpenAI's Training Data — Pratyush Maini, Datology

• The "Reviewer Bottleneck" and AI Fatigue. As AI coding agents (like Codex) generate upwards of 50-75% of code, the human role has shifted from writing to reviewing. This has introduced "AI Fatigue," where developers are exhausted by the high volume of "almost correct" code requiring scrutiny. The industry response involves using AI to review AI (agentic swarms) and implementing strict "plan-first" workflows to minimize the volume of bad code generated.

  • Transcript: 'Inside OpenAI's Agentic Browser, Atlas'
  • The Problem with LLMs
  • I work 12h per day with claude code and don't hit any limits

Dissenting Views

• Simple Retrieval vs. Vector Search. While the industry standard for RAG (Retrieval-Augmented Generation) is vector databases, developers building coding agents argue that keyword-based search (FTS) is superior for code. The consensus view is that semantic search handles ambiguity better, but the dissenting view argues that because LLMs can formulate precise technical queries, the exact-match nature of FTS provides higher fidelity for variable names and specific syntax.

  • What I Learned Building a Memory System for My Coding Agent

• Model Cost vs. Efficiency. Newer, cheaper models (like GLM-5) are marketed as "Opus killers" with massive context windows. However, power users report that these cheaper models often lead to "copy-paste loops" where the model fails to fix bugs iteratively. The dissent here is that despite the 10x+ price difference, using the most expensive model (Opus) is actually more cost-effective due to reduced time-to-solution and fewer wasted tokens on failed attempts.

  • GLM-5 just dropped - here's how to run it in Claude Code in one command
  • GLM 5 is out now.
  • I work 12h per day with claude code and don't hit any limits

Read & Act

What to read - The $285 Billion Crash Wall Street Won't Explain Honestly. Here's What Everyone Missed. — A critical analysis of why the "per-seat" license model is doomed and how to pivot a startup's business model toward "agentic architecture." - What I Learned Building a Memory System for My Coding Agent — A highly technical breakdown of agent memory layers that challenges standard RAG implementations, offering a practical blueprint for building better coding assistants. - Compound Engineering Explained — Defines a new workflow where 50% of engineering time is spent on the "meta-layer" (planning and context optimization) rather than coding, essential for scaling agent use.

What to do - Implement "Core Memory" files: Add a Claude.md or context-pack to your repository root. Use this to store architecture decisions, database schemas, and API patterns so your agents have persistent context without needing full-repo retrieval every session. - Audit your Pricing Strategy: If your startup relies on charging per user, immediately begin modeling "outcome-based" or "usage-based" pricing tiers. Design your product to treat the AI agent as the primary user of your internal API. - Adopt "Plan-First" Coding: Enforce a workflow where the AI agent must generate a detailed plan (and you review it) before it generates code. This "measure twice, cut once" approach reduces hallucinations and "AI fatigue" from reviewing buggy implementations.