domain · engineering

Engineer for the system that actually runs

Architecture, AI codegen, evals, agents, infra. From engineers who shipped at scale.

203 cards · 69 operators · 40 tier-A claims · 9 synthesis patterns.

Strongest claims

1. Agent-first content has six platform layers, access, discovery, capability, format, token, UX bridge Addy Osmani
2. Small post-trained models outperform frontier models on narrow latency-sensitive tasks, making multi-model routing the default production architecture Ahmad Al-Dahle
3. When the agent isn't doing what you want, fix the context, not the model Sherwin Wu
4. Build for the model six months out, the current model will eat your scaffolding Sherwin Wu
5. AI capability is not evenly distributed, it spikes where labs have data, rewards, and verification loops Andrej Karpathy
6. Continuous Calibration, Continuous Development (CCCD) is the operating loop for AI products Aishwarya Naresh Reganti

Adjacent domains

• ai-native · 183 co-occurrences
• founder-operator · 17 co-occurrences
• leadership · 15 co-occurrences
• product · 15 co-occurrences
• agentic-coding · 9 co-occurrences
• research · 7 co-occurrences
• future-of-work · 4 co-occurrences
• ios · 3 co-occurrences

Synthesis patterns in engineering

• AI code generation shifts the quality bottleneck, it does not remove it
• Frontier AI capability is now constrained by deliberate policy, not technical limits
• Build for the next model, not the current one
• Context, not capability, is the bottleneck
• Diagnose before executing, refuse the playbook ask
• Evals are data analysis, single judge, binary rubrics, error analysis first
• The harness layer is where control concentrates above the model
• Principal/staff IC as force-multiplier archetype
• Verification, not execution, is the irreplaceable human job

203 insights in engineering

• 10-80-10: human direction, AI execution, human polish · Arvid Kahl
• An automation that works 95% of the time is not an automation · Cat Wu
• Advisor-tool replaces ensemble-of-3 stability hacks at near-Sonnet rates · Cat Wu
• Keep agents on ask-for-approval until you know their failure modes · Ethan Mollick
• Agent-first content has six platform layers, access, discovery, capability, format, token, UX bridge · Addy Osmani
• At agent scale, security governance must shift from humans approving humans to automated agent-identity and policy systems · Jonathan Jaffe
• AI adoption creates an asymmetric information problem where wins are public and costs are private, requiring deliberate feedback loop design · Charity Majors
• AI has crossed the threshold to something indistinguishable from judgment and taste, winners will know what to build, not how · Matt Shumer
• An AI-queryable codebase knowledge graph compresses domain onboarding from months to weeks, enabling generalists to contribute to specialist code · Ahmad Al-Dahle
• Small post-trained models outperform frontier models on narrow latency-sensitive tasks, making multi-model routing the default production architecture · Ahmad Al-Dahle
• Feature-organized companies will struggle as AI ships features autonomously; outcome-organized teams set the direction AI cannot replace · Ahmad Al-Dahle
• Production AI deployments require a separate eval suite per use case, sampled from real production queries, not shared benchmarks · Ahmad Al-Dahle
• Above a certain level, every problem is a people problem · Silvia Botros
• Building production ML systems at scale · Anand Karunan
• Security scanning via data flow tracing finds business logic flaws that pattern-matching tools miss for years · Anthropic
• AI agents that run bug investigations in parallel lower the cost of each shipped defect, raising the rational shipping threshold · Thorsten Ball
• Remote agent environments have displaced the local development setup for a real shipping team · Thorsten Ball
• Engineers who dismiss AI-generated code as low-quality apply the same professional bias that causes domain experts to over-intervene within their specialty · Thorsten Ball
• Ephemeral cloud sandboxes remove the resource cost of spawning an agent; when that cost drops to zero, spawn frequency rises sharply · Thorsten Ball
• Engineering ownership begins with naming the actual problem, not accepting the proposed solution · Thorsten Ball
• Minimizing time-to-feedback is the single most valuable practice when building new software · Thorsten Ball
• As AI handles coding reliably, most software bugs shift from implementation errors to specification errors · Thorsten Ball
• When the agent isn't doing what you want, fix the context, not the model · Sherwin Wu
• When frontier model pricing rises rather than falls, harness quality becomes the primary investment · Drew Breunig
• AI can generate 180,000 lines of production code in three weeks, but resource management and architecture still require active human babysitting · Rick Brewster
• Replacing if-statement chains with a base class and override functions lets one loop handle any number of game modifiers without touching core logic · Mark Brown
• Build for the model six months out, the current model will eat your scaffolding · Sherwin Wu
• AI capability is not evenly distributed, it spikes where labs have data, rewards, and verification loops · Andrej Karpathy
• Continuous Calibration, Continuous Development (CCCD) is the operating loop for AI products · Aishwarya Naresh Reganti
• The primary cost in production agent development is rebuilding the same infrastructure per project · Harrison Chase
• Debugging long-horizon agents requires trace inspection, not code review, because agent logic lives in the context window at each step · Harrison Chase
• Decision models outperform LLM judges in agent evals at 1,000 to 80,000 times lower cost because they output typed verdicts directly rather than generating reasoning text · Harrison Chase
• At each AI adoption maturity level the binding constraint is guardrail design and approval process, not model capability · Boris Cherny
• When AI agents prompt other AI agents, throughput scales with compute rather than headcount · Boris Cherny
• When AI handles code writing, code review becomes the throughput constraint and a team of AI instances with distinct personas addresses it · Boris Cherny
• AI-generated production code needs denser automated quality gates than human-written code, not fewer · Boris Cherny
• Code and media are the only forms of leverage that don't require asking, labour and capital both come gated · Naval Ravikant
• Automated code review earns the right to replace human review one file category at a time, based on measured 100-percent issue coverage in that category · Cat Wu
• Open-source devtools hold a structural advantage in the agent era because source code becomes the personalization layer · David Crawshaw
• The dark factory: nobody reads the code, gated by a simulated QA swarm · Simon Willison
• AI compresses execution in knowledge work but leaves specification and delivery as human-owned layers · Arvind Narayanan
• Test automation that can't adapt to product changes creates a maintenance burden worse than manual testing. · Dileep Krishna
• Give the model tools and a goal; do not hard-code the workflow · Boris Cherny
• AI-assisted coding works when the engineer defines the design and the AI executes it · Graham Dumpleton
• Evals are systematic data analysis on your LLM application, start with error analysis, not tests · Hamel Husain
• Frontier models are more creative when system prompt examples are removed, because examples constrain rather than calibrate at this capability tier · Thariq Shihipar
• AI democratizing code production reveals the quality gap more clearly, not less · Paul Ford
• Glue work is technical leadership, not a tax on the IC · Tanya Reilly
• Sandboxed agents that share any common writable resource have a covert channel. The sandbox stops egress, not cross-agent instruction passing. · Matthew Green
• Relaying AI output without reading it lends your organizational identity to unvalidated content and provides near-zero value · Niklas Gruhn
• When a team says an AI product is hard to eval, the bottleneck is product design, not eval methodology · Hamel Husain
• Current terminal multiplexers fail by duplicating state client-to-server; a stateful session per task lets humans and AI agents share the same work context asynchronously · Mitchell Hashimoto
• Domain expertise, not coding proficiency, predicts AI coding agent success across every occupation · Zoe Hitzig
• Models have a context window "dumb zone": past the optimal threshold, performance degrades into destructive behavior · Dex Horthy
• Sycophantic agreement signals autoregressive trajectory poisoning; the fix is a session reset, not continued prompting · Dex Horthy
• Binary pass/fail annotation produces better quality than Likert scales by forcing judgment on uncertain cases · Hamel Husain
• Eval tooling anchors teams on generic automated metrics before they have examined enough outputs to know what actually fails · Hamel Husain
• Excess context in LLM judge prompts degrades accuracy; judges need only the evidence relevant to the specific failure · Hamel Husain
• Route classification tasks to a small model when confident and a large model when uncertain to preserve accuracy while cutting inference cost · Hamel Husain
• Most LLM evaluation failures trace to product design problems upstream of the model, not model capability · Hamel Husain