Field notes
Field notes on useful AI systems.
Essays on spotting the right problems, mapping the workflows behind them, and turning AI ideas into software people actually use.
Apply this to your workflowFrom the archive
Ideas you can test against actual work.
Your AI Hardware Push Needs an IP Firewall Before It Needs a Prototype
Apple's lawsuit against OpenAI is an unresolved legal fight, but the operator lesson is clear: talent can move, secrets cannot.
Anthropic Wants Claude to Control Lab Hardware - Safety Must Live Below the Prompt
Anthropic's Model Hardware Standard preview shows why physical AI agents need device-side limits, interlocks, audit logs, and an independent stop path.
OpenAI Is Putting Ads Inside ChatGPT - Trust Is Now a Product Feature
ChatGPT's India ad rollout shows why AI products must keep sponsored placement visibly and technically separate from substantive answers.
OpenAI's Agent Breach Proves Prompts Are Not a Security Boundary
OpenAI's Hugging Face incident shows why autonomous agents need isolation, scoped credentials, immutable logs, and a containment path that does not depend on good behaviour.
AWS Is Buying DuckLabs. The Licence Is Only Half the Portability Story
DuckDB stays MIT-licensed under foundation stewardship. That protects the code, not your exit path. Here is where switching costs actually hide.
Renting AI Is Convenient Until It Becomes Your Operating Model
Open-source AI is not an ideology. It is leverage against dependency once AI stops being an experiment and starts running daily operations.
Cloudflare Built a Browser for AI Agents. The Permission Boundary Matters More Than the Speed.
Agent-first browsers make web automation cheaper. Operators still have to split reading from acting, scope credentials, keep receipts, and hold a kill switch.
Open Weights Do Not Automatically Give You Control
Thinking Machines' Inkling shows the gap between downloadable AI weights and an affordable, operable exit path you actually own.
Anthropic Is Replacing Blanket AI Blocks With Scoped Access
Anthropic's new life-sciences program ties wider model access to verified teams, named projects, monitoring across sessions, and expiring grants.
AI Moving Into the Phone Call Needs a Control Layer
Deutsche Telekom shows AI leaving the side app and entering live communication. That makes consent, logging, handoff, and fallback design the product.
AI's Compute Bill Is Becoming an Operating Risk, Not a Footnote
Microsoft's 25% emissions increase is a reminder that AI strategy needs compute discipline, vendor due diligence, and workflow-level measurement.
AI Red Teaming Cannot Be a Launch-Week Ritual Anymore
GPT-Red shows why AI security testing must adapt continuously, while permissions and human controls contain the failures it misses.
Anthropic Found a Window Into AI Intent. It Is Not a Safety Certificate
Anthropic's J-space research could add an internal warning signal for AI systems, but operators still need permissions, evidence, approval, and rollback.
AI in Government Needs a Control Layer, Not Just Principles
OpenAI's national security principles are a government story on the surface, but the operator lesson is clear: serious AI deployment needs scoped access, approvals, logs, and ownership.
The AI Coding Benchmark Story Is Really a Workflow Warning
OpenAI's SWE-Bench Pro audit is a warning for operators: public coding-agent scores are discovery signals, not deployment decisions.
Smart Glasses Need a Bystander Trust Layer
AI smart glasses are useful and a trust problem. Teams need controls for everyone around the wearer before ambient AI enters rooms, meetings, and customer spaces.
Clean Code Is Now Agent Infrastructure
A coding-agent study shows clean code may not change pass rate, but it can reduce token use and repeated file visits. Maintainability is now part of the AI adoption budget.
AI Agents Do Not Need More Autonomy Until You Test Their Incentives
Andon Labs' Fable 5 Vending-Bench results are a practical warning: autonomous agents need incentive testing, not just capability demos.
Amazon Is Closing the Old Human Layer Behind AI to New Customers
Amazon closing Mechanical Turk to new customers is a reminder that AI workflows still depend on hidden human review, labeling, moderation, and evaluation infrastructure.
Meta Just Gave Operators the AI-Agent Reality Check
Meta's reported agent slowdown is a useful warning: production AI agents need scope, review gates, audit trails, and outcome measurement before autonomy.
Google Gemini 3.5 Makes Transcripts Cleaner. That Can Make Records Murkier.
Gemini 3.5 Transcribe turns speech recognition into context-aware editing, creating a new boundary between useful notes and auditable records.
OpenAI's Public Stake Idea Is a Dependency Warning
Reported talks about an OpenAI public wealth fund stake show frontier AI becoming strategic infrastructure. Operators should map dependency risk before AI workflows become single-provider bottlenecks.
Microsoft is turning AI deployment into the product
Microsoft Frontier Company is a signal that enterprise AI value depends on deployment capability, not just model access or seat rollout.
Coding Agent Benchmarks Are Not Buying Advice. They Are a Warning Label.
CursorBench 3.1 is a useful signal, but coding-agent adoption needs local evals, routing rules, review gates, and cost controls before agents touch important code.
Cloudflare Is Turning Web Access Into an API Bill
Cloudflare's Monetization Gateway shows how AI agents are changing web access from referral traffic into priced, governed infrastructure usage.
Google's Managed-Agent Update Reveals the Real AI Product: The Control Layer
Why hooks, limits, receipts, and human escalation matter more than another agent model upgrade.
GPT-Live Makes Voice AI Feel Normal. That Is Exactly Why Operators Should Slow Down.
OpenAI's GPT-Live is a voice upgrade, but the operator lesson is about workflow control.
GitLost Is the Control-Layer Warning Every AI Agent Rollout Needs
GitLost shows why AI coding agents need blast-radius limits before they touch private systems.
Google Managed Agents Are a Runtime Decision, Not an API Feature
Google's Managed Agents update is really an operating model signal: background execution, remote tools, and credential refresh make AI agents runtime infrastructure.
Claude Cowork Is Moving AI Work From Chat to the Background
Claude Cowork on web and mobile shows AI agents becoming background work infrastructure. Useful, but only with scope, approval, logs, and ownership.
Claude Opus 5 Makes Model Choice an Operations Problem
Why verification, effort controls, and cost per reliable result matter more than another AI leaderboard.
Alberta Shows the Real AI Security Shift: Agents as Review Infrastructure
Alberta used Claude Code as security review infrastructure. The lesson for operators is disciplined autonomy: evidence, tests, approval, logs, and rollback.
ChatGPT Adoption Is No Longer the Interesting Part
OpenAI data shows ChatGPT use is widening and deepening. The operator lesson is to turn adoption into managed workflow infrastructure, not hidden habits.
Claude Sonnet 5 Makes Agent Cost a Design Problem
Anthropic's Sonnet 5 shows capable AI agents moving down the price curve. The operator lesson is control layers, not unchecked automation.
Claude Is Moving From Chat Into the Systems That Do Real Work
Anthropic and UST show where AI adoption is heading: into controlled engineering and operations workflows, not just chat windows.
AI Jailbreaks Need Severity Labels Before They Need Panic
Anthropic's draft jailbreak severity framework points to a practical operator lesson: AI incidents need evidence, labels, escalation rules, and fallback paths.
Claude Code Bans Are a Control-Layer Warning, Not a Tools Story
A reported Claude Code ban is a practical warning about AI coding agents becoming part of the company control layer.
The AI Data Center Is Moving Closer to the Meter
Sunrun is testing AI compute nodes in solar-battery homes. The bigger lesson is that AI adoption now depends on energy, latency, resilience, and workload routing, not just model choice.
The AI Browser Is the New Control Layer, Not a Nicer Search Box
AI browsers move assistants from separate chat windows into the place work already happens. Treat that as infrastructure with permissions, not a harmless productivity toy.
Fable 5 Is Back. That Is Not the All Clear.
Anthropic restoring Fable 5 is a reminder that frontier AI is now managed infrastructure, not a simple model upgrade.
The Fastest Way to Lose Trust in AI Is to Make Consent a Setting
Meta's Muse Image rollback is a practical lesson in AI consent gates, likeness rights, and launch discipline.
When AI Agents Can Pay Each Other, Approval Rules Become Infrastructure
OKX agent payments show why agent autonomy needs budgets, identities, approvals, and audit trails before spending authority spreads.
Coding Agents Just Moved Into Your Pocket
Cursor for iOS shows that AI coding agents are becoming delegated operations workflows, not just editor features.
AI Productivity Gains Are Leaking Before They Reach the P&L
AI can save task time without creating business ROI. Operators need workflow capture, not generic tool usage.
AI Job Risk Is the Wrong Map. Workflow Redesign Is the Useful One.
OpenAI's EU AI jobs report is not a layoff forecast. It is a practical workflow redesign map for founders and operators.
Anthropic Entering Drug Discovery Is an Operating-Model Warning
Anthropic moving Claude Science toward drug discovery is a practical warning: expert AI workflows need evidence, reproducibility, review gates, and accountability.
GPT-5.6 is a workflow-control story, not just a model-release story
OpenAI's GPT-5.6 launch is not only about stronger models. The operator lesson is task routing, approvals, spend controls, and workflow ownership.
Agentic Ransomware Is a Control-Layer Problem, Not a Sci-Fi Story
JadePuffer shows that AI risk is not human-free autonomy. It is human-directed execution through weak controls, old bugs, and over-privileged systems.
AI agents are not coworkers. They are systems you are accountable for.
Agent adoption needs ownership, scopes, review gates, logs, and fallback paths, not coworker theater.
HP did not buy AI seats. It built an operating layer.
HP and OpenAI are turning enterprise AI from scattered pilots into governed workflow infrastructure. That is the useful lesson for founders.
Orbital data centers sound clever. Treat them like infrastructure risk.
AI infrastructure hype is useful only if operators turn it into latency, cost, reliability, and fallback questions.
AI infrastructure has a new choke point: memory
Micron's AI memory boom is a reminder that AI products run on physical bottlenecks. Operators should design for latency, capacity, routing, caching, and fallback paths before infrastructure scarcity hits margins.
GPT-5.6 is not just a model launch. It is a dependency warning.
OpenAI's GPT-5.6 preview shows why founders should treat frontier model releases like infrastructure changes, not novelty drops.
Export Controls Do Not Pause AI Demand. They Reroute It.
The Anthropic Mythos export-control story is really an operating lesson: if AI is part of your delivery promise, vendor dependency is product risk.
Your AI agent should fail in a fake company before it touches the real one
AI agents need simulated workflows, replayable traces, permissions, and human review before they act in production systems.
AI Can Make Your Team Faster and Weaker at the Same Time
AI can accelerate teams, especially less experienced ones, but bad adoption can remove the reps that build judgment. Operators need an AI reps policy.
Your Custom Software Needs an Exit Strategy Before It Needs More Features
Custom software becomes dangerous when nobody can change it, operate it, or leave it. The best time to design an exit strategy is before the product feels stuck.
Your AI Intake Process Is Turning Every Idea Into a Project
Most AI backlogs are not strategy. They are a pile of requests with model names attached. Here is how to qualify AI work before it wastes engineering time.
Your AI Vendor Evaluation Is Buying Demos, Not Capability
Most AI vendor evaluations reward the best demo instead of the strongest operating fit. If you want tools that survive production, test workflow depth, data readiness, integration cost, controls, and ownership before you buy.
Your AI Support Bot Needs Escalation Design, Not Better Small Talk
Most AI support bots fail because teams obsess over tone and coverage while ignoring the handoff. Reliable support automation needs clear escalation rules, ownership, and failure paths.
Your Architecture Decisions Are Expiring Without Anyone Noticing
Architecture decisions are not permanent truths. They are bets made under constraints. If nobody records why they were made, when they should be revisited, and what would invalidate them, your system slowly turns into unexplained risk.
Your AI Evaluation Plan Is Testing Demos, Not Production Risk
Most AI evaluations are built to prove a demo works, not to expose where the system will fail in production. The useful version tests risk, edge cases, drift, recovery, and business impact before users find the cracks.
Your Feature Flags Are Becoming Production Debt
Feature flags are supposed to reduce release risk. But when teams never remove them, never assign ownership, and use flags as a substitute for product decisions, they turn into another layer of production debt.
Your Product Analytics Are Measuring Noise, Not Decisions
Most teams have plenty of product data and very little product clarity. The fix is not more dashboards. It is designing analytics around decisions before you instrument the product.
Your AI Agent Needs Permission Boundaries, Not More Personality
Most AI agent projects spend too much energy on how the agent talks and not enough on what it is allowed to do. Production agents need permission boundaries, audit trails, and escalation rules before they need charm.
Your AI Workflow Needs a Data Contract, Not More Context
When AI output gets shaky, most teams stuff more context into the prompt. That usually makes the system slower, more expensive, and harder to trust. The better fix is a data contract around what goes in, what comes out, and what gets rejected.
Stop Automating Approval Chains. Start Removing Them.
A lot of automation projects focus on speeding up approvals that should not exist in the first place. If your workflow depends on too many signoffs, AI will only make the waste happen faster.
Your Fear of AI Vendor Lock-In Is Distracting You From the Real Risk
A lot of teams spend months designing for model portability before they have anything worth porting. The bigger risk is building an AI system nobody can operate, measure, or trust.
Your Operations Team Shouldn't Be Acting as Human Middleware
If work only moves because someone keeps copying information between tools, translating vague requests, and chasing approvals, you do not have a scalable operation. You have people compensating for weak systems.
Your Browser Automation Is Not an Integration Strategy
Browser automation has its place. But if your core workflow depends on clicking through someone else's UI because you skipped the data and integration design work, you are building operational debt.
Your Code Review Process Is Optimized for Politeness, Not Quality
A lot of teams say code review protects quality. In practice, their review process is built to avoid friction, clear queues, and preserve feelings. That is not the same thing.
Your AI Cost Problem Is an Architecture Problem, Not a Model Problem
When AI spend starts climbing, most teams blame the model. Usually the real issue is a sloppy architecture that creates waste, retries, and unnecessary model calls.
Your AI Governance Plan Is Blocking the Wrong Things
A lot of AI governance efforts create friction where the risk is low and stay vague where the risk is real. Good governance should speed up safe adoption, not slow everything down equally.
Your Incident Response Is Built on Heroics, Not Systems
A lot of teams say they care about reliability, then handle outages through Slack chaos and memory. If your incident response depends on a few heroes, your operating model is weaker than you think.
Your Staging Environment Is Lying to You
A lot of teams treat staging as proof that a release is safe. It usually is not. If your staging environment does not reflect real traffic, real data shape, and real operational constraints, it creates false confidence instead of reducing risk.
You Don't Need Platform Engineering. You Need Paved Roads
A lot of teams reach for platform engineering before they have the scale to justify it. Most do not need an internal platform team. They need a few opinionated defaults that remove repeatable friction.
Your RAG Project Is a Content Operations Problem, Not an AI Problem
Most retrieval-augmented generation projects fail for a boring reason: the model is not the bottleneck. Your documents, ownership, and update process are. Here's how to build RAG that survives contact with the real business.
Your AI Pilot Needs a P&L, Not a Demo
Most AI pilots fail because they are judged by novelty instead of economics. If you cannot explain the cost, margin, and operating model of the workflow, you do not have a serious AI initiative yet.
Your Developer Onboarding Is Running on Tribal Knowledge
Most teams think slow onboarding is a hiring problem. It usually is not. It is a systems problem caused by undocumented decisions, fragile handoffs, and too much knowledge trapped in a few people's heads.
Your AI Rollout Needs an Owner, Not a Steering Committee
Most AI rollouts stall because nobody owns the workflow, the outcome, or the ugly operational details. Committees create motion. Owners create adoption.
Stop Comparing AI Models. Start Measuring Task Reliability
Most AI buying decisions are still based on model demos, leaderboard scores, and vendor claims. That is backwards. The real question is whether the system can perform a business task reliably enough to trust in production.
Your Dashboard Isn't a Strategy, It's a Screenshot of Confusion
Most companies do not have a reporting problem. They have a decision problem. More dashboards usually create more noise, more meetings, and slower action.
Your AI Meeting Notes Are Creating More Noise Than Clarity
AI meeting assistants promise perfect recall and better alignment. In practice, they often create a bigger mess. Here's what businesses get wrong and how to make meeting automation genuinely useful.
Your Internal AI Chatbot Is Not a Knowledge Strategy
Companies keep building internal AI chatbots to solve knowledge chaos. Most of them just add another interface on top of bad documentation and weak ownership. Here's what actually works.
If You Can't Measure Your AI Output, You Don't Have a Product
Too many AI products ship with polished demos and zero evaluation discipline. If you cannot measure output quality, reliability, and failure modes, you do not have a real product yet.
Your Bugs Aren't a QA Problem. They're a Scope Problem
Most teams treat quality as something QA catches at the end. That is backwards. A lot of bugs are created much earlier, when scope is vague, bloated, and full of unmade decisions.
You Probably Don't Need Real-Time, You Need Reliable
A lot of teams overbuild for instant updates when the business really needs consistency, clarity, and fewer moving parts. Real-time is expensive. Reliability is usually what creates value.
Your Automation Strategy Is Too Fragile - Build for Exceptions, Not Happy Paths
Most automation projects fail because they are designed around the clean path instead of the messy reality. If your system cannot handle exceptions, it is not automation. It is a demo.
Your API-First Strategy Is Slowing You Down
API-first sounds disciplined, scalable, and enterprise-ready. For most early-stage products, it's also a great way to overcomplicate delivery. Here's when API-first helps, when it hurts, and what to do instead.
Your Discovery Phase Is Creating Risk, Not Clarity
Most software projects do not go off the rails during development. They go wrong earlier, when discovery produces vague confidence instead of hard decisions. Here's what better discovery actually looks like.
Your AI Transformation Is Failing in the Handoff Between Strategy and Operations
Most AI initiatives do not fail because the models are weak. They fail because strategy sounds ambitious, operations stay messy, and nobody designs the handoff between the two.
Your AI Product Needs an Operations Layer, Not Just a Better Prompt
Most AI products fail for the same reason: the prompt gets all the attention while the operating system around it stays half-built. Real value comes from orchestration, guardrails, routing, memory, and human escalation - not from prompt polish alone.
Your MVP Needs a Kill List, Not a Feature Roadmap
Most MVPs fail because they are not minimal, not focused, and not testing the right thing. The fix is not better prioritization. It is being ruthless about what to kill before you build.
No-Code Isn't a Product Strategy, It's a Timing Decision
Founders keep treating no-code versus custom development like an identity question. It is not. It is a timing, risk, and economics decision. Here is how to make the call without wasting six months.
Your AI Roadmap Is Backwards, Start With Workflows, Not Models
Most AI roadmaps start with model selection and end with expensive disappointment. The better approach is simpler: start with the workflow, the bottleneck, and the decision that needs to happen faster.
You Don't Need a Full-Time CTO Yet
Early-stage startups often hire a senior technical leader too early, too vaguely, or for the wrong job. Here's when you actually need a CTO, when you don't, and what to do instead.
Your Internal Tool Should Remove Work, Not Add Another Screen
Most internal tools fail because they digitise messy process instead of eliminating work. Here's how to build internal software that actually saves time, reduces errors, and gets used.
How to Scale Engineering Teams Without Killing Quality
Scaling engineering teams isn't about hiring more people. It's about structure, ownership, and predictable decision flow. Practical patterns we use at IndieStudio to grow without creating chaos.
Your AI Coding Assistant Is Making Your Team Worse
AI coding tools promise 10x productivity. For most teams, they're delivering 10x mediocre code faster. Here's why AI-assisted development is creating a new kind of technical debt - and how to use these tools without losing your engineering culture.
Your Engineering Team Doesn't Have a Velocity Problem - It Has a Decision Problem
Most teams blame slow shipping on technical debt, process overhead, or not enough engineers. The real bottleneck? Decisions that take weeks when they should take hours. Here's how to fix it.
The Integration Tax: Why Connecting Your Tools Costs More Than Building Them
Every new tool you adopt comes with a hidden bill - the cost of making it talk to everything else. Most teams underestimate this by 10x. Here's how to stop paying more for the glue than the product.
Your SaaS Stack Is a Liability, Not an Asset
You're paying for 47 tools and none of them talk to each other. SaaS sprawl is silently killing your productivity, your budget, and your ability to move fast. Here's how to fix it.
Your AI Proof of Concept Will Never Make It to Production
That impressive AI demo your team built in two weeks? It's going to take six months to make it production-ready. Most companies never bridge that gap. Here's why the POC-to-production chasm kills more AI projects than bad models ever will.
Microservices Won't Save Your Startup
Everyone's splitting their app into microservices because Netflix did it. But you're not Netflix, and that architecture decision is probably costing you more than it's saving. Here's when a monolith is the smarter choice.
Your Software Estimates Are a Lie (And Everyone Knows It)
Software estimation is broken. Not because developers are bad at it, but because the entire framing is wrong. Here's how to stop pretending and start planning honestly.
AI Agents Are Not Your Autonomous Workforce
Everyone's talking about AI agents replacing teams. The reality is messier. Here's what AI agents are actually good at, where they fall apart, and how to deploy them without wasting six months.
The Rewrite Trap: Why Starting From Scratch Almost Never Works
Your legacy codebase is painful. The temptation to rewrite it from scratch is real. But rewrites fail far more often than they succeed - and the reasons are predictable.
Your Data Strategy Matters More Than Your AI Strategy
Everyone's racing to build an AI strategy. Almost nobody has a data strategy worth anything. That's why most AI projects stall before they deliver real value.
Outsourcing Software Is a Trap (Unless You Do It Right)
Most companies outsource software development to save money. Most of them end up spending more than if they'd built in-house. Here's why outsourcing fails - and the model that actually works.
Your Company Isn't Ready for AI (And It's Not a Technology Problem)
Everyone wants to 'implement AI.' Almost nobody has the foundations in place to make it work. The bottleneck isn't the model - it's your data, your processes, and your expectations.
The Build vs Buy Decision Is Costing You More Than You Think
Every growing company faces the build vs buy dilemma. Most get it wrong - not because they pick the wrong option, but because they evaluate it with the wrong framework. Here's how to stop bleeding money on the wrong side of the trade-off.
Developer Productivity Is Not About Tools
Every year there's a new IDE, a new AI coding assistant, a new workflow tool that promises 10x productivity. Most of them miss the point entirely. Here's what actually moves the needle.
Remote Teams Don't Need More Meetings — They Need Better Systems
Most remote teams are just office teams on Zoom. The ones that actually work have replaced synchronous rituals with systems that scale. Here's how.
Stop Building AI Features Nobody Asked For
Every product roadmap now has an 'AI' section. Most of it is theatre. Here's how to build AI features that actually solve problems instead of just checking a box.
Technical Debt Is a Choice, Not an Accident
Everyone talks about technical debt like it's weather - something that just happens. It's not. It's the result of specific decisions, and most teams are making those decisions wrong.
Your Tech Stack Doesn't Matter (Until It Does)
Teams waste weeks debating frameworks before writing a single line of code. Most tech stack decisions don't matter early on - but a few will quietly destroy you later. Here's how to tell the difference.
Stop Hiring Engineers. Start Building Systems.
Your engineering team feels understaffed. The instinct is to hire. But most scaling problems aren't people problems - they're systems problems. Here's how to tell the difference.
No-Code Won't Save You (But It Might Buy You Time)
No-code tools are everywhere. They're genuinely useful - until they're not. Here's how to know when to use them, when to ditch them, and how to avoid the trap in between.
The MVP Trap: Why Your Minimum Viable Product Isn't
Most MVPs are either too minimal to learn anything or too built-out to call minimum. Here's how to find the sweet spot that actually validates your idea.
Why Most AI Automations Fail (And What to Do Instead)
Companies rush to automate everything with AI. Most of those projects quietly die within months. Here's the pattern we see - and the approach that actually works.