Notes from production.
Lessons from systems we run for real clients: what broke, what held, and what we would do again. Written by the engineers who were on call, not a content team.
Your Knowledge Base Is Not Ready for RAG
The model is fine. Your wiki is a landfill.
Read →
Webhooks break quietly. Design for the day they do.
A webhook you don't verify is a promise you can't check.
Read →
Voice Agents Fail on Names, Numbers and Noise
The demo call is quiet. The real call has an impact wrench running behind it.
Read →
The board wants AI. Give them a system, not a demo.
A mandate produces a demo by default. Push it toward a system instead.
Read →
The Agent Loop Is Mostly Error Handling
The loop is 50 lines. Production is everything around it.
Read →
Task limits are how no-code tools tell you to leave
A metered platform charges you more the harder your automation works. That is not a budget problem.
Read →
RPA Clicks Buttons. Agents Make Decisions.
RPA replays a recording. An agent holds a goal and decides.
Read →
OpenAI or Anthropic is a portability question
The real cost is not which model you pick. It is what switching costs later.
Read →
No-shows are a software problem
A missed booking is rarely a flaky customer. It is a confirmation loop with a hole in it.
Read →
Multi-Tenant AI Needs Isolation You Can Prove
A prompt telling the model to stay in its lane is not an access control.
Read →
Multi-Location Businesses Need One Phone Brain, Not Five
Five branches, five phones, five versions of the truth.
Read →
Multi-agent is an architecture tax. Pay it late.
Every agent you add multiplies your failure modes, not your output.
Read →
MCP puts your data one tool call from the model. Act like it.
Every MCP server you connect is a capability you handed a language model.
Read →
Hallucinations are a system design problem, not a model bug
Waiting for a model that never hallucinates is not a plan.
Read →
Guardrails Are Code, Not System Prompts
A system prompt telling the model what not to do is a wish, not a rule.
Read →
Context is an engineering budget
Bigger windows did not end context engineering. They moved where the decisions happen.
Read →
Bad RAG Answers Start in the Ingest Pipeline
The wrong answer shipped weeks before anyone typed the prompt.
Read →
An automation that needs a babysitter is not done
If someone re-runs it every Monday, you built a job description.
Read →
An AI Receptionist Needs an Escalation Plan Before a Voice
Write the list of calls it must never take before you pick a voice.
Read →
AI belongs inside the workflow, not instead of it
Replace the whole process with AI and you lose the parts that worked.
Read →
Agents Earn Trust One Tool Call at a Time
An agent is only as reliable as its riskiest tool call.
Read →
Agentic RAG Is RAG That Knows When to Look Again
Classic RAG retrieves once and hopes. Agentic RAG checks its work first.
Read →
Agentic Coding Pays Off Under Code Review
An agent writes code at the speed your review process allows through.
Read →
A Voice Agent Without Call Review Is Unaccountable
If you cannot read what your agent said yesterday, you are running calls on faith.
Read →
A Model Upgrade Is a Breaking Change
The model behind your agent moves on the vendor's schedule, not yours.
Read →
Your vibe-coded app is not 80% done
The demo working is the first 20%, not the last.
Read →
You did not outgrow automation. You outgrew Zapier.
The ceiling you're hitting belongs to the tool, not the idea.
Read →
A Claude developer is a systems engineer, not a prompt writer
The prompt is the smallest file in the repository.
Read →
Vapi, Retell, Bland: The Platform Is Not the Product
The platform choice is the smallest decision in a voice AI build.
Read →
The premium audit backlog is a data problem
The backlog is not a headcount problem. It is three copies of one file.
Read →
Everyone quotes MIT's 95%. The other finding matters more.
MIT's real finding wasn't the failure rate. It was who avoids it.
Read →
The engineer you interview should be the engineer who joins
Senior in the interview, junior on the invoice, is a business model.
Read →
Structured output is the contract between model and product
Free text is a demo. A schema the rest of your stack can check is a product.
Read →
Sensors do not manage sites. Systems do.
The sensor tells you the tank level. It does not tell you who fixes it.
Read →
An AI agent is production-ready when failure is boring
Readiness is not what the agent gets right. It is what happens when it gets something wrong.
Read →
Outbound AI calls work when the job is narrow
Voice AI fails at persuasion and succeeds at errands.
Read →
Off-the-shelf tools quit before your business does
The tool that got you started was built for someone else's business.
Read →
If You Cannot Run an Eval, You Do Not Have an Agent
A system you cannot test against fixed cases is a demo with opinions.
Read →
n8n is a codebase. Treat it like one.
A workflow that runs your operations needs a rollback plan.
Read →
Meeting notes are where decisions die
A transcript is storage. It is not the same thing as a task.
Read →
MCP turns every integration into the same integration
MCP swaps N times M custom adapters for N plus M.
Read →
Human-in-the-Loop Is a Feature, Not an Apology
The approval step is not a workaround for a weak model. It is the design.
Read →
Compliance belongs in the schedule, not the audit
Checking eligibility after the shift is filled finds nothing but a record of who broke the rule.
Read →
Clinical AI ships when a nurse signs off
Compliance is an accountability design, not a checkbox.
Read →
Build on the Agent SDK when the loop is the product
Pick the framework by where the agent sits, not by GitHub stars.
Read →
Bubble is a real platform when you engineer on it
It fails as magic. It holds as a platform you engineer on.
Read →
Automations fail silently, and that is the expensive part
A stopped automation doesn't error. It just goes quiet.
Read →
An MCP server is an API with opinions
Wrapping an API in MCP is easy. Making a model use it well is the job.
Read →
An AI Receptionist Is a Systems Build, Not a Subscription
A voice model can talk. Turning it into a receptionist is the build.
Read →
A Receptionist That Cannot Book Is an Answering Machine
Picking up the phone is easy. Writing a correct booking is the job.
Read →
