What exactly do you mean by an agent?

Give an agent a job in your business. Give it a clear brief, approved tools and a requirement to record its work.

Think about bringing a new colleague into your firm. You explain the job. You give them access to the information and systems they need. You ask them to write down what they found, what they did and why.

You also explain the things an experienced member of the team would already know. What is normal? What needs attention? What can they decide for themselves, and when must they ask you?

That is a useful way to understand a Sasha agent. It finds answers to business questions and can act within the tools and limits you give it. It can run on an agreed schedule, report what it finds, evaluate what the answer means and recommend what to do. It only does what it is approved to do. Some actions can be approved in advance; others need a fresh decision from a person.

Start with the question you need answered

You do not need to start by designing an agent. Start with a question you keep asking.

• Finance: Where is our money going? What changed? Which payments failed?
• Marketing: Which channels bring enquiries? Which bring sales? What does each cost?

The Finance Agent and Marketing Agent are available now. Other jobs are in development or planned. A Project Agent, for example, would answer: what is late, what is at risk and who needs to do something next?

The question gives the agent a purpose. The brief explains how to do the job.

1. A brief: the job and what good looks like

A brief sets out what the agent should answer, which information to use, how to judge it and when to report. It names the people responsible for following up. It also explains the agent's responsibilities: what it may do within its approved access, the limits it must observe and when it must ask for approval.

It also contains your business knowledge. A new colleague cannot know whether a bill is unusually high until someone explains what you normally spend. An agent needs the same context.

Take Northwind Advisory, an example consultancy with two products: Advisory and Training. Its finance brief might say:

• Review cloud spending each week and split costs between Advisory and Training.
• Compare spending with each product's agreed budget.
• Flag spending more than 20% above budget. This is Northwind's example rule, not a rule for every firm.
• Explain any known reason for the increase. Do not assume an increase is a mistake.
• Give any follow-up action to the person who owns that product.

The answer is more useful because “normal” is written down. A higher bill during a planned Training launch may be expected. The same increase with no known explanation deserves a check.

You can read the brief and correct it. These written instructions are what people mean by prompts. If the budget changes, the instructions can change too.

2. Approved tools: access to do the job

A colleague needs access to the right systems. So does an agent. Northwind's Finance Agent needs to read its cloud bills. Its Marketing Agent needs the connected analytics and sales records that can answer its questions.

An admin approves the tools and access. The starting point is read access, limited to what the job needs. Sasha holds the passwords and keys; the agent uses an approved tool without being able to read or copy those credentials.

Writing “check the bills” in a brief does not create access. The connection must be set up and approved. The brief then explains which accounts, products and reporting periods to use.

Responsibility comes with limits

Just as you would with a colleague, you can approve some work in advance. That approval applies to a defined task, using particular tools and within agreed limits. Having access to a tool does not give the agent permission to do everything that tool allows.

For example, Northwind might authorise an agent to create an internal review task when spending crosses the agreed threshold, where a suitable task tool is connected. That would not authorise it to cancel a service or pay a bill. Anything outside its approved responsibility waits for a person's decision. If approval is unclear, it asks.

The brief describes these responsibilities. Sasha's tool permissions and approval controls enforce the access and action limits. The agent cannot give itself more authority.

3. A work record: what happened and why

You want to be able to go back and inspect the work. Which sources were checked? What did the agent find? What did it recommend? What did you correct?

That record also helps you improve the brief.

Suppose Northwind's agent flags a rise in cloud spending. The owner explains that Training has launched a new service and approves a higher budget. The useful next step is to update the brief with that agreed budget. Future reviews can judge spending against the right expectation.

The record shows why the instructions changed. It also helps distinguish a new problem from an unresolved one.

Today, each agent keeps a record of its runs. The fuller history of individual findings, decisions, reasons and feedback is planned. That fuller history is intended to support comparisons and follow-up across runs.

Improvement comes from inspecting the work and choosing better instructions. The agent does not get to rewrite its own job or permissions.

A report tells you the number. An evaluation explains what it means.

“Cloud spending rose” is a finding. Whether that is acceptable depends on the budget, the reason and what the business got in return.

Website traffic works the same way. Imagine Northwind's traffic rises sharply. Its Marketing Agent checks the connected sales records. The extra visitors mostly read one article, while a smaller referral channel brings Training sales.

If Northwind's aim is sales, more visits alone do not answer the question. The brief tells the agent which results matter and how to judge them. The report shows the evidence and any gaps in it. A recommendation might be to give referrals more attention.

Scripts calculate the numbers and scores. The AI explains and classifies the results using the written rules. A calculation can still use incomplete information, so missing sources must remain visible. A figure the agent could not read says “not read”, never zero.

Three levels of work

1. Report. Find the facts and show the sources: “Training's cloud spending is above its agreed budget.”
2. Recommend. Explain what deserves attention and who should check it: “Ask Training's owner to review the increase against the launch plan.”
3. Act within its authority. Where a suitable action tool is available, carry out work that has been approved in advance and stays within its limits. Otherwise, prepare the proposed action and ask a person to approve it.

The job and its approved tools determine which levels are available. The Finance Agent currently reads and reports; it does not pay bills or change services. A draft message is suggested wording for a person to review and use.

Under the covers, prompts are written instructions

You may hear the word prompt. A prompt is simply an instruction given to an AI model: what to do and how to do it. The model is sometimes called an LLM, short for large language model. It is the AI that reads the information and produces a response.

The brief supplies those instructions. For Northwind, they might say: “Read this week's cloud bills. Compare Advisory and Training with their agreed budgets. Explain anything outside the limits. Show the sources and name who should follow up.”

Writing a better brief means giving clearer instructions, just as you would when explaining a job to a colleague. You can change what the agent looks for, how it judges the answer and how it presents the result.

The brief is kept in a plain document. The descriptions of available tools are written down too, alongside the connections set up to make them work. Sasha supplies the relevant instructions and business information to the model as it works, and the agent keeps records of its runs.

Prompts guide the work. The connected tools let the agent read information or carry out approved actions. Sasha's permission controls enforce what it can access and change. Writing an instruction does not grant new access or approval.

These three parts have separate jobs. You can improve the brief, change an approved connection or review the records without rebuilding everything. The AI model can also be changed without rewriting the whole agent, with checks before adopting the change.

You can inspect the instructions and the recorded reasons for an answer. The agent can still misunderstand something. Sources and records give you a way to check and correct it.

What Sasha adds: a consistent way to organise the job

Agent Creator is opinionated about how an agent should work. It sets standards for the brief, the tools and the record, then fills them with the context Sasha holds about your firm.

• The answer comes first.
• There are at most three recommended actions, each with an owner.
• Every figure has a source. Missing information is labelled.
• Calculations and scoring rules are written down.
• Access starts with reading. Actions need approval, either in advance within defined limits or for the specific action.
• The reporting schedule is agreed. Routine updates stay short; alerts are reserved for urgent matters.

Sasha uses your products, systems, people, responsibilities and working language to draft the brief. You confirm what is right and fill the gaps. Knowing which tools your firm uses helps Sasha propose the right connections; an admin still approves access.

A visible reading list of methods and sources, with suggestions for additions that you choose, is planned. See How agents learn.

Give it a job you can describe

“Each week, tell me what we spent on Advisory and Training, what is outside the agreed budget and who needs to look at it.”

That is a starting brief. Add approved tools, clear limits, a record of the work and your knowledge of how the business runs. You have an agent with a clear purpose, defined responsibilities and work you can review.

See what Agent Creator does, or read about making your own agent.