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, Macro Business, Translation, Application Memory, and LLM Evals.

Key Insights

• Cost Management via "Hill-Climbing" Evals: Early-stage startups face explosive inference costs—one founder reported a jump from $200/month to a $100,000 annualized run rate. The recommended mitigation strategy involves migrating to open-source models by using historical task data to create automated testing loops, allowing engineers to "hill-climb to parity" with expensive proprietary models at a fraction of the cost.
• Will I Be Paid in Tokens?

• Will OpenAI Tank OpenClaw? | E2251

• Agentic Workflow Best Practices: Effective coding agents require a "Junior/Senior" architecture or rigorous CI/CD integration. One successful workflow uses Claude Opus for "feature building" (80-90% complete) and GPT-4/Codex for "review and edge cases." Another approach involves agents generating Pull Requests that must pass automated CI integration tests before human review, ensuring the human remains in the loop for high-level "plan mode" rather than implementation details.

• 🎙️ This week on How I AI: Opus vs. Codex showdown, and AI for accessibility

• Show & Tell: I want to see your Claude code Setup

• Context Engineering for Accuracy: To mitigate hallucinations and generic summaries, prompt engineering must go beyond simple instructions. Techniques include defining explicit "quote rules" (e.g., start/end points, citations) and verifying quotes exist before use. Furthermore, preventing "consensus-seeking" generic answers requires loading four specific context types: Project Context, Business Goal, Product Context, and Participant Overview.

• How to do AI analysis you can actually trust

• Markdown as Application Memory: For application memory and retrieval, Markdown (.md) files are emerging as the standard for "long-term memory" in agents. Because Markdown is lightweight and structured, it allows agents to scan "skill files" and "soul files" (behavior/values) to avoid repetition and build cumulative knowledge, effectively serving as a human-readable database for the AI.

• Is OpenClaw Flawed? Experts React | This Week in AI E001

• Will OpenAI Tank OpenClaw? | E2251

• Metric Shifts in Evaluation: The industry is moving away from generic benchmarks toward business-centric metrics. For CFOs, the new metric is "productive work per dollar of inference," while developers are urged to ignore public benchmarks ("just numbers") in favor of running their own evaluation loops where AI agents grade the output of other AI agents on specific, relevant workloads.

• Will I Be Paid in Tokens?
• Sonnet 4.6 is now the default - do you think this reduces the need for Opus?
• Will OpenAI Tank OpenClaw? | E2251

Emerging Patterns

External State vs. Context Window Developers are increasingly offloading context management from the prompt window to external tools. Rather than passing massive context strings, multi-agent systems are coordinating via "Model Context Protocol" (MCP) to read/write to shared sources of truth like Figma, databases, or issue trackers (e.g., GitHub issues). This keeps the context window clean and reduces "mental fatigue" for the model. - built this video with Claude + Remotion + Figma MCP - Show & Tell: I want to see your Claude code Setup

The "Opus vs. Sonnet" Role Definition A clear bifurcation in model usage is forming. "Opus" class models are being retained for heavy lifting, reasoning, and complex coding tasks where quality is paramount. "Sonnet" class models (specifically 4.6) are viewed as cost-efficiency engines for scaled API products or smaller batches, though some developers report Sonnet 4.6 "over-engineers" responses compared to Opus. - Sonnet 4.6 is now the default - do you think this reduces the need for Opus? - Claude Sonnet 4.6 reaching max usage faster - 🎙️ This week on How I AI: Opus vs. Codex showdown, and AI for accessibility

Dissenting Views

The Productivity Disconnect While individual developers report massive productivity gains (e.g., one engineer shipping 44 PRs and 93k lines of code in 5 days), macro-level data from 6,000 CEOs indicates that 90% of firms see no impact on employment or productivity over the last three years. This suggests a "J-curve" effect where productivity gains are currently siloed in individual high-performers or tech-native firms and have not yet appeared in aggregate economic data. - 🎙️ This week on How I AI: Opus vs. Codex showdown, and AI for accessibility - Thousands of CEOs just admitted AI had no impact on employment or productivity

Read & Act

What to read

• Will I Be Paid in Tokens? — Essential reading for understanding the economics of running an AI startup. It provides a concrete framework (and warning) about inference cost scaling and how to mitigate it via testing loops.
• How to do AI analysis you can actually trust — A practical guide for your "Context Engineering" interest. It moves beyond high-level theory into specific prompting structures to avoid hallucinations and generic "consensus" answers.
• Show & Tell: I want to see your Claude code Setup — A high-signal technical breakdown of a functioning agentic workflow. It details specific tools (VS Code, CI pipelines, issue trackers) that you can immediately replicate.

What to do

• Audit your inference efficiency. Calculate your current "productive work per dollar of inference." If costs are rising, implement a "hill-climbing" test loop: use successful logs from premium models (like Opus) to benchmark and tune cheaper open-source models for specific tasks.
• Implement "Quote Rules" in your prompts. For any retrieval or summarization task, explicitly define the syntax of a valid quote in your system prompt to force the model to verify source existence, reducing hallucinations.
• Adopt the "Junior/Senior" coding model. Experiment with using a high-reasoning model (e.g., Opus/GPT-4) to generate feature plans and code, and a separate, faster model to act as the "reviewer" or "QA" agent to catch edge cases before human review.