5.1Copilot agents: the new generation
The essentials in 30 seconds
- Assistant vs agent: the assistant answers in the moment; the agent is a specialized, persistent colleague, already trained for a role.
- The trio: instructions (role, tone, rules) + knowledge (dedicated sources) + actions (whether it acts or not) — created once, it serves a thousand times.
- Not magic: its quality follows what you give it — and the more autonomous it is, the more its supervision matters.
Running case — Hicham, administrative assistant in Algiers
Hicham, an administrative assistant in Algiers, answers the same questions from his colleagues every day: leave, certificates, internal procedures. Everyone asks him, everyone waits — and he repeats the same answers. He realizes that a specialized Copilot agent, fed with the department's documents, could handle these recurring questions continuously, with no re-briefing. This section tells his light-bulb moment — and the thread continues in 5.2, 5.3, 5.4 and 5.5.
This same case returns in the lesson, the diagram, the script and the 48h challenge — and continues in 5.2, 5.3, 5.4 and 5.5.
1. From assistant to agent
Until now, Copilot was an assistant: you ask, it answers, in the moment. An agent changes its nature: it is a specialized, persistent and sometimes autonomous Copilot for a specific task or domain. It has a role, its own knowledge, sometimes the ability to act. Hicham sees the difference between answering case by case and installing a colleague trained once and for all.
The right image: the assistant is a versatile temp worker you brief every time; the agent is a specialized colleague already trained for a position — the HR agent, the support agent, the compliance agent.
2. What defines an agent
Three attributes distinguish an agent. Instructions (its role, its tone, its rules); knowledge (dedicated document sources, its own knowledge base); and possible actions (querying a system, triggering a flow). Some agents only answer, others act. Hicham already notes the three boxes to fill in for his future administrative agent.
Persistence is the real leap: an agent exists continuously, reusable by a whole team, with no re-briefing. You create it once, it serves a thousand times — that is what makes it pay off.
3. Not magic, but configuration
A sobering guardrail: an agent is not magic. Its quality depends entirely on what you give it — good instructions, clean and up-to-date knowledge, well-framed actions. 'Garbage in, garbage out' matters more than ever. Hicham understands that his agent will only be as good as the department's documents.
And human oversight does not disappear: the more autonomous an agent is (able to act), the more its design, its permissions and its supervision matter. An agent that acts badly causes damage at scale — hence the following sections on careful creation.
Did you know?
Permanent re-briefing is the assistant; the filled position is the agent.
🔎 Concrete example — Hicham turns his repeated answers into an agent
Versatile generalist versus specialized colleague.
Assistant: each colleague rewrites every time a prompt like 'explain the leave policy'.
Hicham's agent: an agent trained on the department's documents answers everyone, continuously, in the department's tone, with no re-briefing.Reading: the agent is created once and serves the whole team a thousand times.
Lesson: its quality depends on clean, up-to-date, well-framed knowledge.
Section vocabulary
- Assistant
- The classic Copilot that answers in the moment, briefed on every request.
Simple example: Hicham prompts each leave question one by one.
What it is not: a colleague trained once who serves with no re-briefing.
Reflex: do I have to re-brief every time?
- Agent
- A specialized, persistent and sometimes autonomous Copilot for a specific domain.
Simple example: Hicham's future agent answers the department's questions continuously.
What it is not: a generalist assistant that knows a little about everything.
Reflex: does this position deserve a permanent specialist?
- Instructions
- The agent's permanent role, tone and rules.
Simple example: Hicham writes: you answer about leave, otherwise you hand off to a human.
What it is not: a throwaway prompt rewritten on every request.
Reflex: are the scope and the handoff written down?
- Knowledge
- The dedicated sources on which the agent grounds its answers.
Simple example: Hicham picks the department's up-to-date documents.
What it is not: general statements with no verifiable source.
Reflex: are my sources authoritative and up to date?
- Persistence
- The agent exists continuously and is reused with no re-briefing.
Simple example: Created once by Hicham, the agent serves the whole department.
What it is not: an answer that disappears after each request.
Reflex: does this repeated task justify a permanent agent?
In practice — Spotting a good agent candidate in your work
- Spot a question or task your team repeats endlessly (leave, certificates, a procedure, a product FAQ).
- Describe the three attributes the agent would have: instructions (role, tone), knowledge (which sources), actions (does it act?).
- Judge the payoff: the more a task is repeated by several people, the more justified the agent is.
- Note the desired level of autonomy: a simple answering agent, or an agent able to act (more supervision required).
Ready-to-say script — proposing a first agent
'In the department in Algiers, I answer the same questions every day — leave, certificates, procedures. I suggest we create a specialized Copilot agent: framed instructions with a handoff to a human outside the scope, knowledge limited to our up-to-date documents, and to start no write actions. We create it once, it will serve the whole team.'
To say to your manager to get a first answering agent approved, with no risky action.
Copy-paste template — agent scoping sheet
AGENT SCOPING SHEET (1 page)
NAME: ___ Agent (e.g. Department leave agent)
REPETITION: repeated question/task ___ · by ___ people · ___ times/week
INSTRUCTIONS: role ___ · tone ___ · DOES NOT DO ___ · handoff to ___ if out of scope
KNOWLEDGE: sources ___ (up to date? ___ / permissioned? ___)
ACTIONS: answers only / acts: ___ (confirmed? ___)
SUCCESS: ___ % of level-1 questions with no human interventionSolo / small budget
Hicham alone: a paper scoping sheet for his most repeated task is enough — no tool, just the three attributes described.
Small business / small team
A single agent candidate for the whole team: the question everyone asks, scoped together in one meeting.
SME in Algiers
Hicham surveys the repeated questions of two pilot departments in Algiers before choosing the first agent to create.
Team getting started
Pair up: one lists the repeated tasks, the other checks that a documented source exists for each.
✓ Best practices
- Reserve the agent for tasks repeated by several people: that is where persistence pays off.
- Think about knowledge first: an agent is only as good as the quality and freshness of its sources.
- The more autonomous an agent is, the more you invest in its design, its permissions and its supervision.
✗ Common pitfalls
- Believing an agent is magic: its quality entirely follows its instructions and its knowledge.
- Creating an agent for a one-off task: without repetition, persistence brings nothing.
- Giving autonomy with no supervision: an agent that acts badly causes damage at scale.
Key takeaways
- The assistant answers in the moment; the agent is specialized, persistent and sometimes autonomous.
- Three attributes: instructions (role), knowledge (dedicated sources), actions (whether it acts or not).
- Persistence is the real leap: created once, reused by the team a thousand times.
- An agent is not magic: its quality follows its data, and the more it acts, the more it must be supervised.
48h challenge — apply tomorrow
Fill in the scoping sheet above for 1 repeated task of Hicham's: do the three attributes fit on one page?
- I can explain assistant vs agent in one sentence.
- My agent case is scoped (role, sources, does it act?).
- I know why a poorly scoped agent answers badly.
Frequently asked questions
What is the concrete difference between an assistant and an agent?
The assistant is briefed on every request; the agent is a Copilot already trained for a role, persistent, reusable by a whole team with no re-briefing.
Can all agents act on systems?
No: some only answer from knowledge, others have actions (querying, triggering). The more an agent acts, the more its design and supervision matter.
What distinguishes an agent from a simple Copilot assistant?