Agentic Coding

Briefing: Agentic Coding Purpose: I'm a software engineer who's building with AI coding agents daily. Tell me what's actually working and what isn't. Looking for practical tips I can apply.

Key Insights

• The engineering role has shifted from authoring to pipeline management and verification. Engineers are spending significantly more time on upfront specification and reviewing massive, complex PRs generated by agents. To cope with the sheer volume of AI-generated code, enterprise teams are building sophisticated pre-processors and "critic engines" to grade agent PRs, filtering out low-value noise and surfacing only high-confidence changes for human review.
• From Software Engineer to Agent Manager - How Work is Changing in A New Software Development Paradigm
• Uber: Leading engineering through an agentic shift - The Pragmatic Summit

• From Coder to Manager: Navigating the Shift to Agentic Engineering with Notion Co-Founder Simon Last

• Agent architectures are converging on a Planner-Worker-Judge hierarchy. Top AI labs and tools (including Cursor, Anthropic, and OpenAI) have independently abandoned "flat" multi-agent structures because agents sharing a single file become risk-averse and collide. The successful pattern is decomposition, parallel execution in isolated sandboxes, and verification by a distinct "judge" agent that can cleanly restart the context window if a task fails.

• 4 AI Labs Built the Same System Without Talking to Each Other (And Nobody's Discussing Why)

• Terminal workflows are beating IDEs for asynchronous "toil" tasks. While IDEs are great for synchronous generation, engineers are increasingly using CLI tools (Claude Code, Codex, Gemini CLI) combined with tmux and SSH to spin up background agents on remote sandboxes. At companies like Uber, 70% of agentic workloads are asynchronous "toil tasks" like library upgrades, migrations, and unit test generation, allowing developers to act as tech leads directing parallel workers.

• State of Agentic Coding #4 with Armin and Ben
• Uber: Leading engineering through an agentic shift - The Pragmatic Summit

• From Coder to Manager: Navigating the Shift to Agentic Engineering with Notion Co-Founder Simon Last

• Model selection is highly task-dependent, with GPT-4 currently favored for long-horizon execution. While Claude Opus and newer Claude models offer excellent ergonomics, practitioners find that GPT/Codex models consistently outperform Claude on long-horizon, complex agentic tasks because Claude tends to take shortcuts or stop early. Cost is also a major factor, driving teams to use expensive frontier models purely for planning, while handing execution off to cheaper, smaller models.

• 4 AI Labs Built the Same System Without Talking to Each Other (And Nobody's Discussing Why)
• Uber: Leading engineering through an agentic shift - The Pragmatic Summit

• State of Agentic Coding #4 with Armin and Ben

• Monoliths and human-centric APIs are fatal to agent context windows. Monolithic codebases (anything over 500k to 5M lines) fundamentally break current agentic workflows because they cannot fit into context windows. Furthermore, tools are actively being redesigned for "Agent Experience" (AX)—for example, Notion moved away from verbose JSON to SQLite and custom Markdown specifically because standard API structures waste tokens and confuse agents.

• From IDEs to AI Agents with Steve Yegge
• From Coder to Manager: Navigating the Shift to Agentic Engineering with Notion Co-Founder Simon Last
• Introducing Proof

Emerging Patterns

• The "Vampiric" Burnout and Code Degradation. Despite high output, the developer experience is suffering. Practitioners report a "vampiric effect" where managing parallel agents causes extreme cognitive fatigue from relentless context switching. Simultaneously, software quality is degrading; "vibe-coded" applications are becoming increasingly CPU and memory intensive, and unmonitored agents are injecting "heresies"—incorrect architectural patterns that spread virally through a codebase.
• From IDEs to AI Agents with Steve Yegge
• State of Agentic Coding #4 with Armin and Ben

• From Software Engineer to Agent Manager - How Work is Changing in A New Software Development Paradigm

• Encoding "No" as the New Competitive Moat. Because code generation is effectively commoditized, the most valuable engineering skill is now recognizing when AI output is flawed and articulating why. Organizations are shifting focus from prompt engineering to systematically capturing these rejections and turning them into durable, reusable constraints (a "critic engine"), ensuring agents don't make the same mistakes twice.

• Stop accepting AI output that "looks right." The other 17% is everything and nobody is ready for it.
• Uber: Leading engineering through an agentic shift - The Pragmatic Summit

Dissenting Views

• The Illusion of "100x" Productivity. Consensus narratives highlight massive productivity gains, citing engineers shipping 10 to 30 pull requests a day and operating 30x faster than traditional developers. However, veteran practitioners argue these numbers are highly misleading. A deep dive into GitHub commit spikes reveals that much of this new output consists of disposable side projects, "slop forks" (reimplementing existing apps for fun), and experiments, rather than a true acceleration of complex, core business logic.
• Consensus: The Org Chart Math Behind AI-Native Speed
• Dissent: State of Agentic Coding #4 with Armin and Ben

Read & Act

What to read: - Uber: Leading engineering through an agentic shift - The Pragmatic Summit — Essential viewing for anyone managing agentic code at scale. Breaks down exactly how an enterprise builds infrastructure to handle the noise of AI-generated PRs and orchestrate async migrations. - 4 AI Labs Built the Same System Without Talking to Each Other (And Nobody's Discussing Why) — A concise breakdown of the "Planner-Worker-Judge" architecture. Critical for understanding why flat, single-agent scripts fail in production. - Stop accepting AI output that "looks right." The other 17% is everything and nobody is ready for it. — A vital perspective shift on evaluation. Explains how to build a compounding "taste" engine by formally encoding your rejections of AI output.

What to do: - Separate Planning Models from Execution Models. Stop using your most expensive models (like Opus or GPT-4o) for the entirety of a coding task. Use the frontier models to generate the technical spec and architectural plan, then hand the execution off to faster, cheaper models to manage quadratic token costs. - Move "Toil" out of the IDE. Stop using Cursor for migrations, dependency updates, and boilerplate tests. Shift these asynchronous, well-defined workloads to CLI agents running in remote tmux sessions so you can continue focused, synchronous feature work locally without breaking your context. - Implement a "Critic Engine" in your CI/CD. Do not let agents merge code based solely on green tests. Build an automated reviewer step that explicitly checks agent PRs against a library of your past "rejections" and domain-specific architectural constraints before it ever reaches a human reviewer.