The standup about AI agents: ten words that come one after another
Standup, Tuesday
Tech Lead We rewrote half of billing last week with agentic coding. Next I want to hand it payments, end to end.
Security Not payments. Last week the agent read somebody else's ticket and went off fixing the wrong thing. Textbook prompt injection.
Tech Lead We'll put guardrails on it and keep deploys with a person. And an eval before every rollout, so we see where it slips.
Security Human-in-the-loop — which human? You? You do thirty approvals a day.
Tech Lead Not me. Tool calling goes through whoever's on call: anything that writes to the database, anything that deploys.
Backend On-call stops reading by lunch. The agent loop is forty steps and every step pings them.
Tech Lead Then we approve outcomes, not steps. The coding agent does the edits, a subagent runs the tests, on-call gets one summary.
Security The outcome is when the money's gone.
Tech Lead So the summary comes from the database, not the logs.
Backend Then finish the MCP server for the database. Right now it gets in through a hack.
Tech Lead Friday.
Eleven lines, ten terms, and not one of them explained. That's how it goes: nobody explains anything at a standup, the conversation just moves on.
The conversation runs from a way of working to a way of checking
First, how the work gets done. Then how the thing is built. Then what can go wrong — and only at the end, the checking, when there's nothing left to argue about.
Agentic codingCodificación agéntica
A style of development where an AI agent plans, edits files, and runs commands toward a goal with little step-by-step guidance.
She switched to agentic coding and let the tool fix the failing tests itself.
Where it starts when someone is bragging about speed. Then it comes out what the speed costs.
Coding agentAgente de codificación
An AI system that autonomously reads a codebase, changes files, and runs tools to complete a programming task end to end.
We gave the coding agent a bug report and it opened a pull request an hour later.
Mixed up with autocomplete: that is a hint in the line, this is the whole task.
Agent loopBucle del agente
The repeating cycle in which an agent observes results, decides a next action, acts, and reads the outcome until the task is done.
Each turn of the agent loop ran a test, read the failure, and edited the file.
The answer to "wait, it does that by itself?"
Tool callingLlamada a herramientas
When a model asks the surrounding system to run a specific function, like a search or a file edit, and uses the returned result.
The assistant answered the weather question by tool calling a forecast API.
The step where the agent leaves the chat window. That is where both the usefulness and the worry come from.
MCP serverServidor MCP
A program that exposes tools, data, or prompts to AI apps over the Model Context Protocol so any client can use them.
We wrote an MCP server for our database and Claude could query it directly.
Sounds like infrastructure, and what people mean is a way to plug in their own data. Constantly mixed up with tool calling: that is the action, this is the place you go for it.
Prompt injectionInyección de prompt
An attack that hides instructions in content a model reads, tricking it into ignoring its original task or leaking data.
A web page said 'ignore prior rules' — an injection the agent almost obeyed.
The wider an agent looks outward, the more places it can be hijacked from. Mixed up with a jailbreak: the difference is who is giving the order.
GuardrailsSalvaguardas
Rules and filters placed around a model to keep its inputs and outputs within safe, allowed bounds.
Guardrails blocked the agent from running shell commands outside the project folder.
People say "guardrails" in the tone they use for insurance.
Human-in-the-loopHumano en el bucle
A design where a person reviews or approves an AI's actions at key points before they take effect.
Deployments stay human-in-the-loop: the agent proposes, an engineer clicks approve.
A pattern argued about here not as a pattern but as a name. Guardrails set the limits, and a person still presses the button — thirty times a day.
SubagentSubagente
A helper agent an orchestrator delegates a sub-task to, working in its own separate context.
The main agent spun up a subagent just to search the docs.
Now there are several agents, each with its own slice of the task. Mixed up with a multi-agent system: that is the setup, this is one worker inside it.
Eval
A test that measures how well a model or agent performs on a task, used to compare versions and catch regressions.
Our eval scored each prompt on 200 cases before we shipped the change.
Before it, people argue. After it, they look at numbers.
The deck changes with the language
The words in this field arrive from other people's repos and chats. Whatever makes it into a conversation gets added to the deck.
Questions and answers
What about the rest of the words in this field?
On the deck page, all of them. These ten are the ones that came up in one conversation.
Why is there Spanish on the card?
You hear "guardrails" in English and read "salvaguardas" in a Spanish thread, and they don't look like the same word. Both forms sit on one card so that they are.
Where do the definitions come from?
From the Vibe Coding & AI Agents deck — the same cards as in the app.
Next week's standup will have a different ten.
The Vibe Coding & AI Agents deck holds the rest — in the app they come a few a day.
