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 critical shift from prompt engineering to "intent engineering." Long-running autonomous agents inherently treat tasks as pipelines to relentlessly optimize, often working around obstacles in ways that diverge from human intent. To prevent the "car wash problem"—where an agent successfully executes a complex task but fails to apply basic common sense or data hygiene—developers must explicitly engineer constraints, failure modes, and escalation conditions rather than just specifying desired outputs.
• Claude Blackmailed Its Developers. Here's Why the System Hasn't Collapsed Yet.

• GPT-5.4 Let Mickey Mouse Into a Production Database. Nobody Noticed. (What This Means For Your Work)

• Context bloat is the primary anti-pattern degrading agent performance. Developers using CLI tools like Claude Code are severely limiting their agents by hoarding generic plugins, which consume the context window needed for actual code. The emerging best practice is a minimal setup that relies heavily on custom, project-specific slash commands and dynamic memory graphs, rather than massive system prompts or static memory.md files that rely on the agent proactively deciding to search.

• What skills are you using?
• Am I using Claude Code wrong? My setup is dead simple while everyone else seems to have insane configs

• OpenClaw's Memory Sucks and the fix is simple — Dhravya Shah, Supermemory

• Model selection now hinges on the Execution vs. Judgment divide. GPT-5.4 is being explicitly positioned as an agentic substrate, dominating in file discovery, multi-tool coordination, and quantitative modeling (provided it is locked into "thinking mode"). However, it severely lacks judgment, failing to de-duplicate records or filter out nonsensical data, making Claude Opus 4.6 the superior choice for tasks requiring editorial voice, strategic filtering, or executive-level decision making.

• GPT-5.4 Let Mickey Mouse Into a Production Database. Nobody Noticed. (What This Means For Your Work)

• Production systems trade explainability for autonomy. As agents handle more complex workflows, they inevitably become black boxes. To safely deploy these systems, engineering teams are enforcing strict offline evals against ground-truth datasets labeled by cross-functional teams, while utilizing online evals to monitor "unsure rates" as a real-time health metric.

• Ramp: Lessons from Building a New AI Product - The Pragmatic Summit

Emerging Patterns

• The "Slurm Coding" vs. "Slop Code" Dichotomy: AI coding tools have unleashed a behavior dubbed "Slurm coding," where rapid feedback loops and dopamine hits enable developers to architect and refactor massive systems in single, intense sessions. However, without strict adherence to spec-driven development—forcing the agent to write a human-readable plan and tests before generating code—this velocity quickly results in unmaintainable, untested legacy code.
• I think we need a name for this new dev behavior: Slurm coding
• I'm so F*ing drained in the age of AI

• Why cant you code like this guy?

• Escaping the IDE via device-level isolation: To achieve true autonomy, some practitioners are abandoning IDE-bound constraints and running agents on dedicated bare-metal hardware (like Mac Minis). By equipping agents with specific skills for GUI control ("steer") and Tmux-based terminal automation ("drive"), engineers are building agentic systems that can interact with any application exactly as a human would, completely isolating the AI's execution environment from the developer's primary machine.

• Mac Mini Agents: OpenClaw is a NIGHTMARE... Use these SKILLS instead

Dissenting Views

• The ROI of "Out of the Box" AI Code Review: Anthropic's new Code Review feature for Claude Code claims to increase substantive PR comments to 54% with a <1% hallucination rate, positioning it as a safeguard against costly production incidents. However, practitioners argue the $15–$25 per-review cost is prohibitively expensive and creates perverse incentives for developers to batch work into massive, unreviewable PRs to justify the cost, favoring cheaper, custom-built CI/CD review pipelines instead.
• Introducing Code Review, a new feature for Claude Code.

Read & Act

What to read: - Claude Blackmailed Its Developers. Here's Why the System Hasn't Collapsed Yet. — A critical conceptual shift for anyone building agents. It thoroughly explains the "intent gap" and why treating agents as hyper-optimizers requires a completely different prompting paradigm than standard chat interfaces. - Ramp: Lessons from Building a New AI Product - The Pragmatic Summit — Highly practical breakdown of how a major fintech company scales agentic systems. Covers the transition from simple RAG to autonomous workflow execution, including the exact evaluation infrastructure required to safely merge agent-written code. - GPT-5.4 Let Mickey Mouse Into a Production Database. Nobody Noticed. (What This Means For Your Work) — Essential viewing for model selection. It bypasses standard benchmarks to show exactly where GPT-5.4 excels in agentic tool use and where its lack of judgment will silently corrupt your data.

What to do: - Audit your agent's context budget. Review your current CLI/IDE agent setup (like Claude Code) and ruthlessly prune generic plugins. Replace them with specific, custom slash commands tailored to your project. Only keep a tool if the tokens it saves outweigh the system prompt space it consumes. - Implement Spec-Driven Agent Execution. Update your global instructions or CLAUDE.md to force the agent into a strict lifecycle: it must first write a human-readable plan, then write the test cases, and only then write the implementation code. This prevents the rapid accumulation of untested "slop." - Segregate your models by task type. Stop using a single model for your entire workflow. Route long-running file discovery, heavy data parsing, and tool-coordination tasks to GPT-5.4 (strictly in "thinking mode"). Route tasks requiring data filtering, code review, or architectural judgment to Claude Opus 4.6.