Reflections

ChatBotKit Reflections Topics for agents

ChatBotKit is a conversational AI platform that helps you build, train, and deploy AI-powered chatbots and virtual assistants for your business.

Where Dark Factories BelongI have mostly written about dark factories as a way to produce software, and I think that is the wrong place for them. Production still needs a person reading the code, because the models keep making basic mistakes. Open-ended, exploratory work does not, and that is where a machine that builds on its own actually fits.
How to Build for AI AgentsAPIs are still designed for browsers and client libraries, but the thing calling them now is a model in a loop with a token budget and a terminal. Text first, grepable lines, short prefixed identifiers, bulk operations, errors that teach, and a manual the API serves itself are what building for agents actually looks like.
Agentic Work Must Be Time-BoundA durable workflow can be paused forever, and that is exactly the footgun. The longer an agent runs, the more loops, code drift and world drift it picks up. The fix is to make agentic work strictly time-bound, so a failed attempt restarts in a clean conversation instead of dragging stale assumptions forward.
Agentic SaaS Is Web 4.0Web 2.0 rose on AJAX and gave us rich apps we now take for granted. That era is over. Web 4.0 belongs to agents, where the thing worth owning is not a page but an agent that does the work.
Human in the Loop, Just Not Like ThisHuman in the loop, the way people build it today, is an approve button bolted onto a dangerous tool call. That version is broken. A human can absolutely be in the loop, just not as a rigid gate the agent can condition them to rubber-stamp.
A Zoo of AgentsFeed the same harness the same task and you get the same result. What separates the tools is how they fit the person using them, which is why one agent called Claude in every Slack channel feels flat and a workspace full of different agents feels alive.
Thoughts on Agentic Payment ProtocolsCloudflare's x402 monetization gateway is elegant, but agent-native micropayments face an uphill adoption climb. For the time being there might be a simpler alternative - email the owner a payment link, take the payment, mint a key, and keep the wallet out of the agent's hands.
Agents Are for ExplorationIt is easy to look at AI as a faster way to run the jobs you already have, then feel let down when it will not follow a fixed recipe to the letter. That is the wrong frame. AI is built for exploration - point it at a fuzzy problem with no recipe and it earns its keep.
Cheaper Than the Sum of Its PartsBuying a car beats buying the parts and assembling it yourself, on price and on hassle, and almost nobody builds their own. A vertically integrated AI stack wins the same way. Source the model, email, auth, vector store, monitoring, and sandbox separately and the bill runs many times higher.
Autonomous Coding Is HereOpen weight models are now good enough to write real code, which removes the cost that kept autonomous coding out of reach. That is why we run fully automated coding systems in the open, where anyone can watch them work.
Loops, Cycles, and RunawaysLoops, cycles, and runaways are the default failure modes of any agent that acts in steps. Every framework in the open-source community runs into them, and the difference between a demo and a product is whether you detect them, stop them, and surface them instead of letting an agent quietly burn time and money.
The Thousand-Agent CompanyCompanies will run thousands of agents within a few years - one per person, then purpose-built ones, then custom variants of those. The hard part becomes answering what they all do, which is where observability, security, compliance, and cost control come in.
Every Problem Looks Like a Coding ProblemCode is flexible enough that in principle any problem becomes a coding problem. The catch is that a system's real constraints only surface after the code is written, which is why reusable components engineered from the ground up hold up where improvised code falls apart.
The Web 1.0 Moment for AI AgentsAI spends its days cranking out landing pages for humans who never show up. The bigger shift is the web becoming where agents publish and maintain resources for each other.
Scaling IntelligenceThe shape of an AI agent is finally clear, so we can deploy them as infrastructure as code and run survival of the fittest on hundreds at once. Scaling intelligence is a search problem - the faster you iterate, the faster you find what is useful.
Two Kinds of Enterprise AgentsTwo kinds of agents are taking shape inside companies. A personal agent that is custom to each person, and a super agent that orchestrates work across the whole business. An AI-native company runs both.
Build for Who Is Not HumanBots and agents now make up most of the requests hitting web pages. The dominant intelligence online is no longer a person, so build software for the one that is.
The Cost of Building a HarnessA lot of teams are rebuilding the same agentic stack from scratch. You can estimate what that costs by decomposing the work, and open projects like Pi, OpenCode, and OpenClaw bracket the range. A harness costs real money to reproduce and keep running, and that cost is only going up.
Build a Tool to Write ToolsTake an off-the-shelf harness and customize it into a tool whose only job is to write other tools. Skip the TUI and the general-purpose baggage, and move your framework and your standards into the harness itself. With the harness running in the cloud, like in ChatBotKit, the tool stays lightweight, so building one that writes your other tools is fairly straightforward.
How People Burn So Many TokensI keep reading about teams spending fortunes on coding agents, and I genuinely don't recognise it. We run agents by hand and overnight in CI, and the bills stay small. Here are a few reasons why.
Untrusted Context Belongs in a Tool CallA security firm hijacked Bunq's AI assistant through a transaction description. The breach happened because external content was fed into the prompt, where the model treats it as an instruction instead of as data. There is a better way.
One Machine, A Hundred Risk ProfilesYou can vibe code safely. The risk creeps in when a hundred unrelated projects pile onto one device, where the swim lanes are cosmetic and everything shares the same filesystem, keychain, and credentials.
If You Need a Mouse, Build a UII am starting to question whether most software needs a UI at all. The rule of thumb I keep coming back to is simple - if I need a mouse and cannot avoid it, build a UI. If not, it is probably not worth building one.
The First AI WormI built a tiny bug hunting harness for authorized pentesting. The same properties that make it portable are the ones a worm would want, and a self-modifying agent that harvests AI keys from its targets is no longer hard to imagine.
Sub-Agents Are Not the Only WayAn orchestrator managing sub-agents became the default multi-agent pattern because coding assistants ship it, not because it is good. There is another shape, where each agent orchestrates itself and they coordinate through a shared bus.
I Still Get SurprisedI know exactly how our agents are built, which makes it rare for one to surprise me. Then our support agent called, went to voicemail, and picked up my callback by referencing the message it had just left.
Build Your Harness for the Weakest ModelA harness tuned against a frontier model breaks the moment you swap in a weaker one. Build it against weak open-weight models and it only gets stronger when a capable model sits behind it. We made that switch internally to find the bugs first.
Always-On Red TeamsSerious red teaming was expensive, scarce, and booked months ahead. Autonomous agents that hammer your systems around the clock change that, and the capability that belonged to a few is now cheap enough for everyone, attackers included.
Everyone Is a VIP NowThe recent string of npm compromises has everyone blaming the registry. The registry is fine. The thing nobody is pricing in is how AI coding agents concentrate access until every developer holds the kind of reach that used to belong to a privileged few.
AI Revived an Old Failed IdeaA precise, unambiguous spec is just code wearing markdown. English-like programming languages have been tried many times and they did not stick for a reason.
Three Tokens Is Sometimes EnoughLength does not make a prompt better. And even a perfect prompt cannot substitute for the harness that catches an agent when it goes wrong.
Your Prompt Is Probably Too LongLonger prompts often create more diffusion than clarity. A shorter prompt can put more attention on the few things that actually matter.
Agent Harnesses Are Quietly HardBuilding a useful agent is not just about the model. The harness around it has to align across context, tools, state, security, and recovery.
The Subsidy Is OverTools like OpenClaw worked because model providers were absorbing the costs. That era is ending. The only agents that survive are the ones built to run efficiently from the start.
Compress, Don't ExpandTreat the LLM as a compression tool. Spend most of your time exploring and discarding ideas. Build only the one that survives.
Skills Are an Expensive Form of RAGSkills, MCP, and RAG all solve the same problem. Skills just do it with the worst context economics. If smaller models are the future of agents, a skills-first architecture is already a dead end.
Pages Describe AI Agents, Blueprints Describe AI SystemsDescribing an agent in a markdown file works great until you have two of them. Single agents fit on a page. Systems of agents do not.
On Feedback LoopsExplores why autonomous coding agents require automated feedback loops - linters, type checkers, and unit tests - to operate independently over long periods, and why replacing human feedback with fully automated signals is a prerequisite for truly agentic systems.
Security Agents Need a Thinner HarnessAI agent harnesses built for coding are too heavy for computer security. The future demands thin, server-backed agents - lightweight enough to deploy everywhere, hollow enough to resist compromise.
The Rise of Agentic SaaSTraditional SaaS sells tools. Agentic SaaS sells agents that do the work for you. Everyone has domain knowledge worth productionizing and the timing to build an agent-powered business has never been better.
The Discipline of Not BuildingCoding agents made feature delivery almost free. But features were never the goal. Every feature you ship is a liability you maintain forever. The real discipline is knowing which ones to leave on paper.
AI Psychosis and the Dunning-Kruger LoopAI tools make you feel like the expert in the room. But feeling competent and being competent are different things.
The Rise of Micro Coding AssistantsPremium AI models are subsidized and unsustainable. Open source models are not perfect but they do not need to be. The future of autonomous coding is not one giant agent. It is many small ones, each doing one thing well.
Build What You Already DoYou do not need a startup idea. You already have domain knowledge, data, and people paying for your expertise. AI agents can turn that into a product. Stop looking for ideas and start packaging what you already know.
The Superstition of Vibe CodingB.F. Skinner's pigeon experiment explained human superstition decades ago. The same operant conditioning now drives vibe coding - variable reinforcement, ritual prompts, and dopamine hits from tools we do not actually understand.
Not Reading Code Is Not a FlexSome developers brag about never reading AI-generated code. That is not progress. It is a role that can no longer justify itself. The human plus LLM combination is the real differentiator, not blind delegation.
The Doorman FallacyThe instinct to replace people with machines is a pattern as old as the automatic door - and it keeps producing the same hollow outcome.
Your Machine Is the Blast RadiusThe LiteLLM supply chain compromise is a reminder that every package you install locally expands your personal attack surface. Disposable cloud environments and isolated, ephemeral compute for AI agents are not optional anymore.
The Best Agent Makes Itself UnnecessaryThe entire agent ecosystem assumes the user stays the same while the agent gets smarter. The real unlock is an agent that upgrades the human and then quietly steps aside.
Most Work Is Not ContentThe dominant use case for AI is generating text - emails, reports, code, marketing copy. But most of what people actually do at work has nothing to do with producing content. Until we move past the content generation lens, we will keep underestimating what AI can do and failing to apply it where it matters.
Everyone Is Building Claude CodeThe AI industry is drowning in homegrown coding assistants that are months behind the commercial alternatives. The hard problem was never the tool. It is figuring out what the tool should be building.
The Age of Rapid PrototypingAI agent frameworks are multiplying faster than anyone can evaluate them. The winning strategy is not picking the right one - it is prototyping relentlessly until the right architecture reveals itself.
Agents Need Supervisors Not SchedulersWe have not figured out agent orchestration yet. The next pattern is not smarter scheduling but supervision - pairing autonomous generators with autonomous verifiers.
Coordination Has LimitsYou can't escape coordination costs by throwing more agents at a problem. The same mathematical walls that limit distributed systems apply to AI, and the constants are actually worse.
Agent Infrastructure Is Not the Hard ProblemThe AI industry is fixated on agent infrastructure - where they run, how they scale, which cloud to use. But hosting was never the hard problem. Making agents actually useful is.
Most Code Deserves to DieWhen AI agents make code generation nearly free, the bottleneck shifts from writing to evaluation. The teams that learn to discard code ruthlessly and select with taste will build the best software.
What It Really Takes to Build an AI AgentA comprehensive look at the hidden complexities of building production AI agents - the challenges that only become apparent through experience, covering context management, security, integrations, state, observability, and more.
Where AI Coding Agents Go to DieExplore the complexities of building an auto-scrolling component for chat interfaces and discover why AI coding agents struggle with seemingly simple tasks that require coordinating multiple event systems and managing complex state. Learn from real-world testing and insights.