Job definition
We write down the one job the agent does, the requests it should handle, the ones it hands to a person and what finished looks like. You receive: a one-page job definition, for you to approve.
AI agent development
We build agents that work inside your CRM, calendar, ticketing and email with narrow permissions, a person approving the steps that matter and a log of what they did.
Scope your AI agentWhat an agent is
An AI agent development company builds software that reads a request, decides which steps it needs and carries them out in your CRM, calendar, ticketing tool or inbox. It acts through accounts with narrow permissions, and a person approves the steps that matter.
It answers a question in a chat window and stops there. It changes nothing in your systems, so the risk is a wrong answer, and the work is language and accuracy.
When a form arrives, step one runs, then step two. The path is fixed in advance. That makes it predictable and light to run, and it is the right tool for many jobs.
It reads a messy request and decides whether to look something up, hold a slot, open a ticket or ask a person. Each tool it can call is listed, with what it may and may not do.
Agent or workflow
Both can read a request and update your systems. They differ in who decides the steps, what happens on an odd input and what each costs to run.
| A fixed workflowThe steps are written in advance | An agent with a narrow jobThe model picks the steps | |
|---|---|---|
| Who decides the steps | The builder, in a diagram that runs the same way each time | The model, from a list of tools it may call |
| An unusual request | It follows the closest branch, or fails and alerts someone | It can reason about it, or hand it to a person |
| Cost to run | Few or no model calls, so cost per run is low and steady | Several model calls per request, so cost follows the work |
| How you test it | A few sample inputs, since each path is known | A set of realistic requests, scored on what the agent did |
| Suits | Inputs with a clear shape and steps that rarely change | Inputs that vary a lot and need judgment between steps |
When you can write the steps on one page. A form that creates a CRM contact and sends a confirmation does not need a model choosing what to do, and a workflow costs less to build and to run. We say so in the scope, and quote the workflow. An agent earns its place where the input is open-ended text, the next step depends on what the request says, and a person would otherwise read each one.
What the sources say
Anthropic, OWASP and NIST on what an agent is, where it goes wrong and how to hold it in check. The source sits under each point.
Its engineering guide calls workflows systems where models and tools follow “predefined code paths”, and agents systems where the model “dynamically” directs its own process and tool use. We use the same line when we scope a job: you see which steps are fixed and which the agent decides.
The guide says this “might mean not building agentic systems at all”, and that agentic systems often trade latency and cost for better task performance. Where the job could be done either way, we say so in the scope.
Its 2025 entry on excessive agency gives examples: a read-only task connected through an identity that can also update and delete, or an extension that deletes documents without asking the user. The scope lists every tool the agent can call and the permission each one holds.
The same entry recommends “complete mediation”: your CRM or calendar validates each request against its own rules. It also lists human-in-the-loop approval for high-impact actions. Our agents connect through accounts that hold only the permissions listed, and the approval step is in the written scope.
Its prompt-injection entry describes indirect injection: instructions hidden in content the model reads, such as a web page or a file. It says retrieval and fine-tuning do not fully mitigate it, and lists least privilege and human approval for high-risk actions. A customer’s email is such content, so an agent that acts on it could be steered into a wrong action. The test set includes emails written to try exactly that.
NIST released version 1.0 on January 26, 2023, and a Generative AI Profile, NIST-AI-600-1, on July 26, 2024. NIST describes it as voluntary guidance. We borrow its vocabulary for the scope: who owns the agent, what it is allowed to do, how it is measured and how it is managed once live.
The scope
Six pieces of an agent build, each one a document or setting you can inspect afterwards.
We write down the one job the agent does, the requests it should handle, the ones it hands to a person and what finished looks like. You receive: a one-page job definition, for you to approve.
We list each system the agent touches, each action it can take there and the permission behind it. Reading, drafting and creating are separated from changing and deleting. You receive: an action register, with each permission set in an account in your name.
We name the actions that wait for a person, who approves them, and what the person sees: the request, what the agent proposes and the source it used. You receive: the approval rules and the screen or message used to approve.
We collect real requests from your inbox, write the correct outcome for each, add messages that try to push the agent off its job, and agree the score that counts as ready. You receive: the test set and the pass line, in your files.
We record what the agent read, what it decided, what it did and what it handed over, with the time and the request it came from. You receive: access to the log, kept in a system in your name.
We set a cap on steps per request, a cap on spend and an alert when either is reached, and estimate model cost per request before launch. You receive: the limits sheet and a cost estimate, for you to approve.
How it runs
Nothing starts until you approve the scope, and you approve each stage before the next one. Each stage hands you a document for your files.
Action register
| Action | System | Permission | Approval |
|---|---|---|---|
| Look up customer and past rentals | CRM | Read only | None |
| Check equipment availability | Scheduling | Read only | None |
| Hold a slot for 24 hours | Scheduling | Create a hold only | None |
| Open a damage or complaint ticket | Ticketing | Create only | None |
| Confirm a booking and send the reply | Scheduling and email | Create booking, send email | A person approves |
| Refund, waive a fee or change a price | Billing | No access | Staff only |
Our stack
The models, orchestration and monitoring an agent runs on, and the systems it connects to. Select a tool to see what it is and how it fits an agent build.
Claude is Anthropic’s family of language models. Through its API, a model can ask for one of the functions you list to be run.
We use Claude where the agent reads long or messy requests and chooses between listed tools, and write its data terms into the scope.
Questions
Keep exploring
A short paid review: which AI use cases pay, and which to skip.
Salesforce, HubSpot or Zoho, set up around how your team sells.
AI built on your documents and data, tested before launch.
HubSpot configured, connected and loaded with clean data.
A website chat that answers from your approved content.
An AI phone line that answers, books and hands off.
Phone lines for reminders, status calls and after hours.
Zoho CRM and Zoho Books set up with clean data.
One repeated process mapped, then automated.
Start here
Tell us where the agent would work. You get a written scope with the price, and you can decide after reading it.