AI
Treat your agent harness like production code, not like a config file
The instructions, tools, and sandboxes that run your AI agents are a software product. Most teams aren't treating them that way, and it shows.
2 min read
Journal
Thoughts on building, shipping, and everything in between. Engineering deep-dives, product stories, and industry perspectives from the studio.
35 posts
- AI
Fine-tuning a small model beats frontier APIs—but only under four specific conditions
Airbnb's production benchmark shows a 2B-parameter fine-tuned model matching GPT-class quality at 35ms and a hundredth of the cost. The conditions that make it work are concrete, and one of them is harder than it looks.
3 min read
- Engineering
The new bottlenecks in AI-era engineering are human, not process
Shipping 11x more code removes old constraints like review queues and communication overhead, but the replacements—thrashing, lost shared context, and cognitive exhaustion—are harder to automate away.
3 min read
- Product
Sell separately when demand outstrips supply — bundling makes it worse
New research from Kellogg School shows that once inventory falls below 30–60% of demand, unbundling extracts more revenue than any package deal. Here's why that matters for software products with real capacity constraints.
3 min read
- AI
Real incidents make better benchmarks than synthetic tasks for AI agents
PostHog's approach to evaluating AI agents—using real past incidents as ground truth and frozen environments for isolation—shows why synthetic evals measure fluency, not usefulness.
3 min read
- Engineering
AI-generated tests should surface signals, not block merges
Meta's JiTTesting system caught 74 serious privacy bugs before they landed in production. The real lesson isn't the AI—it's the deliberate choice to keep humans on the loop, not in it.
2 min read
- Industry
The discipline behind turbopuffer's $100M ARR and 50-person team
turbopuffer built over $100M in annualized run rate with about 50 people and minimal venture capital. The hiring discipline and customer orientation behind that ratio are worth understanding.
4 min read
- Product
Holiday discounts for subscription apps work — but only in the right windows
RevenueCat's data on 5,000 apps shows Black Friday and New Year promotions pay off, Christmas doesn't, and half the gain comes from users already in your free tier.
4 min read
- Product
When everyone can build, who decides what ships?
AI has collapsed the product and UX roles into a single "builder" position at many companies. The real problem isn't the role change—it's that the judgment layer, the one that asks whether something should ship at all, quietly disappears with it.
3 min read
- AI
Context engineering is a permanent discipline, not a configuration problem
Context engineering stays hard because models aren't people—their construction creates failure modes that scale with capability. The same six tools keep working against ever-new problems, and that stability is worth understanding.
3 min read
- AI
AI products rot differently: model drift is a quality problem, not just an ops problem
Traditional software breaks visibly. AI products degrade silently—users stop trusting a feature before they stop using it. That distinction changes how you need to think about quality.
3 min read
- AI
Specialist models beat frontier models when you can define the task precisely
Shopify's approach to fine-tuned subagents is compelling, but the real investment isn't the training—it's building a reliable evaluation harness. That's the part most teams underestimate.
3 min read
- AI
The default 0.50 threshold is quietly breaking your AI evals
Decision models for automated evals are only as good as the threshold you tune them to. Most teams never touch the default—and pay for it in missed failures and false alarms.
3 min read
- Product
The token pricing trap bites builders, not just end users
AI's credit and token pricing models obscure value for end customers, but the confusion flows upstream too—hurting every team that builds on top of AI APIs.
4 min read
- Engineering
Separate your control plane from your data path before you need to
Cloudflare's D1 post shows why routing metadata and lifecycle operations must never touch the same path as live queries—and what happens when you get that separation right from the start.
3 min read
- AI
Building internal AI tools: what Ramp and Stripe's bet reveals for small teams
Ramp and Stripe built their own internal vibe-coding platforms instead of buying vendor tools. The reasons are more instructive than the trend itself.
3 min read
- AI
The bottleneck in AI-generated code is verification, not generation
Giving AI agents a self-check loop matters more than model quality. Here's what that means for teams building with AI today.
3 min read
- Engineering
When your tool's core assumptions flip: lessons from Anthropic's custom VCS
Anthropic replaced git with an in-house system called Scion after AI agents made git's offline-first design a liability. The real lesson isn't about version control—it's about recognizing when a tool's foundational assumptions have inverted.
2 min read
- Product
Predictability is the thing silently killing your product roadmap
Marty Cagan's SVPG piece reframes a familiar debate: the real enemy of good products isn't output, it's the organizational hunger for predictability. Here's what that means for small teams.
3 min read
- AI
AI agents don't fail at SQL—they fail at knowing which data to trust
LinkedIn's data context layer improved acceptable answers from 25.9% to 79.5% by giving its analytics agent structured, curated knowledge about datasets—not by improving the model. The lesson matters for anyone building AI-assisted workflows.
4 min read
- AI
A green build from an AI agent proves less than you think
When an AI agent writes both code and tests, it encodes the same misunderstanding in both. Independent product verification — rooted in user requirements, not agent interpretation — is now a required discipline.
4 min read
- Engineering
Synthetic database benchmarks give you false confidence before migrations
Airbnb's production traffic replay system caught latency regressions that synthetic tests would have missed entirely. Here's what that means for teams who can't build the same infrastructure.
3 min read
- Engineering
Code review is theater now, and quality has to move somewhere else
Agent-authored PRs now outnumber human ones on GitHub, and nobody can review that volume carefully. Quality hasn't disappeared — the signal has just moved.
3 min read
- Engineering
Test your core by building the thing that uses it, then throwing it away
Ephemeral testing uses a disposable layer on top of your base component as the test signal. With AI making throwaway prototypes cheap, it's worth taking seriously.
3 min read
- AI
Model name is not a capability guarantee
A 43,000-call analysis shows that reasoning budget—not model identity—determines what users actually get. For product builders, that changes how to evaluate, contract, and monitor AI components.
4 min read
- Product
When building gets cheap, prioritization becomes strategy
As AI lowers the cost of writing code, the real bottleneck shifts to choosing what to build. James Shore's financial modeling approach makes that choice rigorous instead of political.
3 min read
- Engineering
Row-level security in PostgreSQL can cost nothing or cost seconds — the difference is in the details
RLS performance in PostgreSQL hinges on three specific policy mistakes. The worst part: small tenants in your test data will hide the problem entirely.
3 min read
- Engineering
When code becomes cheap, your data model is the only thing with permanent weight
AI makes features and layouts reversible. That asymmetry doesn't reduce the importance of data decisions—it sharpens it.
3 min read
- AI
Dishonest by design: the verification gap in AI products
Glyph's essay argues that current AI tools disclaim unreliability while designing every interaction to discourage verification. We agree — and the responsibility extends to any small team shipping AI features.
3 min read
- Engineering
Fix the network boundary before tuning your worker pool
WHOOP cut a 15.8-million-task ML pipeline from two months to six days. The biggest win wasn't parallelism—it was removing an HTTP call they didn't need.
3 min read
- Engineering
CPU spot pricing is gone — treat compute like a scarce resource
AI-driven demand has wiped out CPU spot markets and pushed reservation lead times to months. Small teams need to rethink the assumption that compute is always available on demand.
3 min read
- Engineering
Cloud sandboxes solve where agents run, not how well they reason
Docker's new cloud sandboxes handle isolation and always-on compute for long-running coding agents. The harder problem—keeping an agent on task over hours of unmonitored work—still lives above the infrastructure layer.
3 min read
- Product
Shipping a SaaS in 2 weeks: lessons from building Mocha
How we went from idea to paying customers in 14 days — our process for rapid product development with Next.js, Stripe, and BunnyCDN.
3 min read
- Engineering
Dual-pass transcription: how we handle real-time medical AI
Why we run two AI models in sequence — one for speed, one for accuracy — and what we learned about German medical terminology along the way.
5 min read
- Founder
Building for the German healthcare market as a bootstrapped founder
DSGVO, KBV compliance, medical vocabulary, and the reality of selling AI tools to therapists and doctors in the DACH region.
3 min read