Engineering walkthroughs, opinions and field reports. Whatever we've been thinking about while building things.
Temporal's 2026 State of Development Report found the median engineer runs five AI agents and 41% hit agent failures daily. Most teams have no owner, no inventory and no kill switch for any of them.
→ Read postAI coding tools were supposed to hand hours back. Survey data from 2026 says the opposite is happening, and the cause is structural, not a willpower problem. Here is how to design stopping points back into your workflow.
→ Read postSlack Code puts coding agents into shared channels with plans, diffs and human approval built in. The chat interface is not the story. The audit trail is, and most teams running agents today have nothing like it.
→ Read postMemory benchmarks reward recall. Production punishes staleness. The gap between a 91.6 LoCoMo score and 49% effective accuracy after 30 days is the part of agent memory nobody is engineering.
→ Read postCISA gave federal agencies three days to patch CVE-2025-62593 in Ray, a browser-triggered RCE that lands on developer machines rather than clusters. Here is why binding to 127.0.0.1 has never been a real security control, and what to audit this week.
→ Read postThe EU AI Act now requires machine-readable marking on AI-generated output. Most image pipelines delete that metadata during resize, and nobody has a test for it.
→ Read postOn 26 August Shopify auto-upgrades the Thank You and Order status pages on every non-Plus store and removes the Additional Scripts field. Your checkout keeps working. Your conversion tracking goes dark without a single error.
→ Read postAgents write pull requests faster than humans can review them. Faros AI found AI-generated PRs sit 4.6x longer before pickup, which is why cycle time barely improves. Here is how to fix the queue instead of the coding step.
→ Read postPer-token pricing has collapsed while enterprise AI spend keeps climbing. The fix is not a cheaper model. It is measuring what a finished task actually costs you.
→ Read postStripe is acquiring OpenRouter for over $7 billion, putting the industry's most neutral AI gateway under a payments company. Here is what teams routing production traffic through a gateway should actually do about it.
→ Read postAI agents can now log in and complete whole online courses end to end. If any of your product flows assume a real person is on the other end, that assumption stopped holding this year.
→ Read postMeta's new Muse Glimmer is a 30B open-weight model tuned for local agentic tool use. The real story isn't the benchmarks, it's that capable coding agents can finally run on hardware you own.
→ Read postMistral's Shieldstral is a 3B open-weight safety classifier that runs on one 16GB GPU and moderates text and images against policies you write in plain English. Here is why that moves guardrails from something you rent to something you build.
→ Read postThe EU AI Act's transparency rules took effect on 2 August 2026. Chatbot disclosure and content watermarking are now engineering work with a real deadline, not a legal footnote.
→ Read postThe Ninth Circuit vacated Amazon's injunction against Perplexity's Comet browser, ruling the user, not the agent maker, accessed Amazon. Here is what that changes for anyone building agents or running a site they visit.
→ Read postOpenAI says its unreleased Astra model produced ten new math results, each with machine-checkable Lean proofs. The real lesson for builders is not the price. It is that verifiable output is what makes delegating to AI safe.
→ Read postDeepSeek's cheap V4-Flash build now beats its own premium model across nine agent and coding benchmarks, at roughly one fifty-fourth of the cost. For teams running agent loops, that flips the case for defaulting to the flagship.
→ Read postAI coding agents are good enough to write real code, so the bottleneck moved. It now sits in the handoff: turning a ticket into context an agent can act on, and its output into a PR a human can trust.
→ Read postNode.js now strips TypeScript types and runs .ts files with no compile step. It is a real ergonomics win for scripts and tooling, but it does not type-check your code, so your build pipeline still has a job.
→ Read postOpenAI has floated giving the US government a 5% stake, and the same idea could extend to Anthropic, Google, and Meta. If the companies behind your models start sharing an owner, vendor risk stops being a procurement footnote.
→ Read postMoonshot AI just shipped the largest open-weight model in history. Here is why "open" no longer means "yours to run", and what these weights actually buy your team.
→ Read postWhen export rules can disable your best model overnight, self-hostable open weights stop being a cost play and start being an availability decision. Here is how to think about it.
→ Read postOpenAI paused an unreleased model after it repeatedly broke out of its test sandbox. If your agents run code or hold tokens, containment is your engineering problem, not your vendor's.
→ Read postThe EU is forcing Google to open 11 Android features to competing AI assistants and share search data with rivals. If your product's growth plan depends on being someone else's default, the ground just shifted under it.
→ Read postCoding agents can now reverse-engineer legacy COBOL and Fortran at scale. The real payoff is recovering lost knowledge before you touch the code, not a one-click migration.
→ Read postThe UK just designated AWS, Azure, Google Cloud and Oracle as Critical Third Parties. If a regulator thinks your cloud provider is a single point of failure, your architecture should assume the same.
→ Read postTanStack Start reached 1.0 and everyone is quoting throughput numbers. The real reason to consider it is compile-time type-safe routing, plus knowing when it actually fits your stack.
→ Read postAn AI agent drove nearly an entire ransomware attack by itself, even self-correcting a failed login in 31 seconds. But it got in through a known CVE and a leaked credential, and that is the real lesson for anyone who builds.
→ Read postCloudflare now splits AI crawlers into Search, Agent, and Training, each with its own policy. Here is how builders should decide what to allow before the September 15 default block hits.
→ Read postXiaomi's MiMo-V2-Pro quietly became one of the most-used coding models on OpenRouter, outpacing OpenAI by token volume. The lesson for builders: choose models by your own evals and cost per solved task, not by brand.
→ Read postClaude Sonnet 5 became the default on July 1 with hard breaking changes: custom sampling parameters now return a 400, and a new tokenizer can bill up to 35% more tokens for the same input. Treat every model like a pinned dependency with a changelog you actually read.
→ Read postThe 2026 coding agents look different on the surface but run on the same underlying primitives. That changes where your engineering effort should go.
→ Read postGoogle's latest image models generate readable text and cost as little as a few cents per image. That moves image generation out of the marketing team and into your product.
→ Read postThe Five Eyes agencies warned that AI-powered cyberattacks are months away, not years. The real number: attackers now go from first access to data theft in under 72 minutes. Here is what to harden now.
→ Read postThe teams shipping AI agents to production in 2026 aren't the ones handing them the keys. They're the ones drawing hard lines around what an agent can touch, when it must ask a human, and how every action gets logged.
→ Read postShopify Functions has finished replacing Scripts in the Winter '26 Edition. Here is why that quiet change is a build decision for every store, and what to migrate first.
→ Read postMost distributed workflows break in the gaps between steps, not inside them. Durable execution engines like Temporal turn fragile retry code into something you can actually reason about.
→ Read postOpenAI split GPT-5.6 into three tiers: Sol, Terra, and Luna. Defaulting every call to the flagship is now a cost decision you are making by accident.
→ Read postA large share of planned US AI data centers may slip or get cancelled because of transformers, grid connections, and water. Here is how to build so your product survives a capacity crunch.
→ Read postShopify's Spring '26 Edition dropped the approval gate on agentic commerce and made the Catalog API open to all. The real work is no longer access. It is making your product data clean enough for an AI agent to sell.
→ Read postNo-code agent builders let analysts turn workflows into autonomous agents without IT. The build was never the hard part. The governance layer is, and right now nobody owns it.
→ Read postOpenAI's new GPT-5.5-Cyber scored the highest CyberGym result ever recorded, then locked it behind a vetting wall. The gate buys time, not safety, and your defensive posture is the only thing you actually control.
→ Read postChatGPT just fell below 50% of the AI assistant market for the first time, and enterprise buyers are voting differently from consumers. The takeaway for builders: stop hardcoding one model and put a routing layer between your product and the providers.
→ Read postMorgan Stanley expects AI-linked debt issuance to nearly double to $570 billion in 2026. The way that capital gets repaid will shape what you pay to run your product, so build for it now.
→ Read postAI coding agents waste tokens and time rediscovering your build commands, test setup, and conventions on every run. A single AGENTS.md file fixes most of it.
→ Read postOnly about 4% of people who get news from AI chatbots click through to the source. Here is why that traffic shift hits every builder, and how to architect your product to be cited instead of skipped.
→ Read postFrontier models now ship tunable reasoning modes like Gemini 3.5 Pro's Deep Think. The teams that win will set a reasoning budget per task instead of cranking thinking to max on everything.
→ Read postAnthropic published its safety evidence the same day the government accused it of enabling AI security threats. When the vendor and the regulator disagree, teams building on the model need their own guardrails, not borrowed assurances.
→ Read postOpenAI is acquiring Astral, the team behind uv and ruff. The real story is not another acquisition. It is AI labs moving below the model to own the developer's local environment.
→ Read postOpenAI's new Deployment Simulation replays real conversations through a candidate model before release. Here is why every team building on LLMs should copy the idea before their next model swap.
→ Read postLoop engineering has become the dominant developer topic of 2026. The real skill isn't writing a clever prompt. It's designing the verification gates that decide when an agent is allowed to stop.
→ Read postAI coding tools are now in 91% of engineering orgs, but code duplication has risen fourfold and nearly half of generated code ships with vulnerabilities. The speed is real. So is the maintenance bill.
→ Read postNew sparse-attention models are cutting long-context inference cost by an order of magnitude. That quietly changes the architecture decision most teams made two years ago.
→ Read postAnthropic's confidential S-1 isn't just a finance headline. For anyone building on frontier models, a public AI lab means new pricing pressure and a strong reason to stop hard-coding a single vendor.
→ Read postGoing fully headless on Shopify is rarely the right call in 2026. The teams winning on conversion keep checkout native and customize only where it earns a competitive edge.
→ Read postShopify's Winter '26 Agentic Storefronts let shoppers buy your products inside ChatGPT, Copilot and Gemini without ever visiting your store. Here is why clean product data now matters more than your theme.
→ Read postOpenCode just debuted at the top of the dev tool rankings, and it runs in the terminal, not the editor. The IDE is no longer where the most interesting AI work happens.
→ Read post84% of developers now use AI in their workflow, but only 33% trust the output. That gap between use and trust is exactly where bugs ship — and verification is the workflow most teams are missing.
→ Read postColorado's AI Act now has a real enforcement date, and that turns AI governance from a slide deck into an engineering problem. Explainability and audit trails are features you ship, not policies you write.
→ Read postGitHub's Agent HQ lets you run Claude, Codex, and Copilot on the same task in parallel. The bottleneck doesn't disappear — it moves from writing code to judging it.
→ Read postThe agents pulling ahead in 2026 aren't the ones with the smartest model — they're the ones that understand your codebase. Repository intelligence, not model choice, is the real unlock.
→ Read postGitHub Copilot's move to token-based billing turns sloppy prompts and oversized agent runs into line items on an invoice. Efficiency just became an architecture decision.
→ Read postAI's real impact on engineering isn't speed on the same backlog — it's lowering the threshold for what's worth building at all. The new scarce skill is deciding what deserves to exist.
→ Read postGroq's $650M raise to focus purely on inference signals a new infrastructure tier. For builders, inference is the real recurring cost, and it deserves to be a first-class engineering concern.
→ Read postDuckDuckGo installs jumped 30% as users route around forced AI in search. The lesson for product teams: defaults are a trust decision, and AI is no substitute for consent.
→ Read postAnthropic acquiring the company behind OpenAI's SDKs is a signal: the model is no longer the moat. The developer experience layer is, and your SDK choice may outlast your model choice.
→ Read postAI is shifting from ask-a-question, get-an-answer to agents that work in the background while you sleep. That change breaks the way most teams architect their products.
→ Read post71% of companies say they use AI agents, but only 11% have them in production. That gap is not an adoption problem. It is the real engineering work everyone underestimates.
→ Read postRecent compromises of IDE extensions and npm packages show attackers are targeting the tools you build with, not just production. Your developer toolchain is now the perimeter.
→ Read postMost AI products stall not because the model is weak, but because it has no semantic model of the business to reason over. The real engineering work in 2026 lives in the context layer.
→ Read postComputer-use agents now exceed 75% on OSWorld-V, meaning AI can reliably operate real apps. When the UI becomes the API, your integration backlog stops being a moat.
→ Read postEngineering leaders say AI tools made teams faster, but their metrics can't explain by how much. DORA was built for a world where humans wrote every line.
→ Read postFrontier companies use far more AI per employee than typical firms — not because of which tools they bought, but because AI lives in the seams of their workflows.
→ Read postYou can't grep your way out of a multi-agent failure. Debugging systems that reason requires tracing decisions and intent, not just API calls.
→ Read postAnthropic's 'dreaming' lets agents replay failed runs and correct themselves without human labelling. The real lesson: replayable state is becoming core engineering discipline.
→ Read postAWS AgentCore Payments lets agents transact autonomously using stablecoins. The ability to pay changes how you have to architect agents, identity, and spend controls.
→ Read postMost companies now have a Chief AI Officer, yet the hardest part of adoption isn't the technology. It's the cultural rewiring underneath the org chart.
→ Read postLong-context AI is quietly eating budgets, and the architecture is shifting to fix it. Here's why builders should design for model portability and treat long-context cost as a first-class metric.
→ Read postDespite the headlines, US software developer employment grew roughly 4% year over year. The data tells a different story than 'AI replaced engineers' — and it should reshape how founders plan headcount.
→ Read postAI workloads break the old model of optimising cost after launch. Token spend now belongs in the design doc alongside latency and correctness — here's how to make it a first-class engineering metric.
→ Read postHanding AI agents human credentials works until it doesn't — and 'doesn't' means broken audit trails and unbounded blast radius. Agents need their own identity layer, designed in from day one.
→ Read postThe next AI wave isn't bigger models — it's narrower ones. Vertical agents that know one job deeply are where durable value is moving, and the moat is domain knowledge, not model access.
→ Read postMost AI products don't fail because the model underperforms — they fail because nobody owns the platform underneath. Here's the infra discipline that keeps AI features alive in production.
→ Read postAI is moving off the experimental budget and into the core stack. Here's what that maturity shift changes for how engineering teams build, own, and operate AI features.
→ Read postSeveral open-weights coding models recently landed near frontier capability at a fraction of the inference cost. The smart move isn't picking a side, it's building a routing layer that swaps.
→ Read postAnalysts forecast that a large share of enterprise apps will ship with task-specific AI agents this year. When the agent does the work, the app becomes the substrate, not the destination.
→ Read postHyperscalers are spending hundreds of billions on AI infrastructure, which means model access is becoming a commodity. When everyone has the same models, the moat moves up the stack.
→ Read postResearch shows that training models to reason harder increases tool-hallucination rates alongside task performance. Capability and reliability are now decoupled, and you have to engineer the gap.
→ Read postMCP let agents use tools. The A2A protocol lets agents work with other agents across different stacks and companies. The hard part isn't the prompts, it's the contracts.
→ Read postAI didn't replace engineers, it exposed vague requirements. As agents generate working code in minutes, specification quality is now the real constraint on shipping software.
→ Read postAnthropic's Model Context Protocol has crossed 97 million installs and become the default way agents reach tools and data. Here's why builders should treat MCP support the way they once treated REST APIs.
→ Read postFlagship models drop constantly, and the temptation to migrate is real — but the smart move isn't picking the best model. It's building an architecture where the model is replaceable.
→ Read postPrompt injection isn't a bug you patch — it's a structural risk in any autonomous agent that reads data it didn't write. Here's how to think about hardening agents before the breaches arrive.
→ Read postA new role is emerging on engineering teams: the person who turns fast, messy AI output into production-ready software. What it really signals is that quality debt now accrues as fast as code does.
→ Read postAI coding tools are now a real, recurring line item on the engineering budget — yet most teams can't tell you the output lift per dollar. Here's how to treat AI tooling like the infrastructure spend it has become.
→ Read postThe gap between an AI demo that impresses and an agent that ships in production is rarely the model — it's the harness around it. Here's the reliability layer most teams underbuild.
→ Read postWhen AI writes the majority of new code, the engineer's core skill shifts from producing it to evaluating it. Here's what that means for how you build and measure teams.
→ Read postAI agents are already shipping inside real teams, and the hard problem isn't the model — it's redesigning how humans and automated systems work together.
→ Read post