Key takeaways
- Microsoft and IBM both describe AI agents as systems that use AI models to plan and take actions with tools towards a goal.
- Agents range from assistants that act only when asked to autonomous agents triggered by events, usually with human approval for risky steps.
- Agents 'go rogue' through bad instructions, excessive permissions, prompt injection or loops; limits and monitoring prevent most incidents.
- Agents improve through better instructions, tools and data, and sometimes fine-tuned models; they have short-term and stored memory by design.
- Many agents are sold as SaaS or through marketplaces such as cloud and app stores.
AI agents, as Microsoft and IBM both describe them, are software systems that use AI models to understand a goal, plan steps and take actions with tools, such as searching data, calling APIs or updating records, with varying degrees of autonomy. Microsoft builds them into Microsoft 365 Copilot and Copilot Studio; IBM offers them through watsonx Orchestrate. This page explains each company's view, how autonomous agents are, how they go wrong, whether they can be trained or remember, what they run on, whether they are SaaS, where to sell them, and when they started.
What are AI agents at Microsoft?
At Microsoft, AI agents are assistants built on Microsoft 365 Copilot and Copilot Studio that work on behalf of a person or team: they answer from organisational knowledge, take actions through connectors and APIs, and can run on triggers or schedules.
Also asked as: what are ai agents microsoft
Microsoft offers prebuilt agents, such as research and analysis agents in Microsoft 365 Copilot and agents in Teams and SharePoint, and lets organisations build their own with Agent Builder for simple cases or Copilot Studio for workflows and actions. Developers can also build agents with Azure AI services and frameworks, and connect tools through the Model Context Protocol. Agents inherit Microsoft 365 permissions and admin controls, which matters for enterprise data.
What are AI agents according to IBM?
IBM describes an AI agent as a system that can autonomously perform tasks on behalf of a user or another system by designing its workflow and using available tools. Its main platform for building and running agents in businesses is watsonx Orchestrate.
Also asked as: what are ai agents ibm
IBM's framing emphasises enterprise use: agents that automate processes in HR, procurement, customer service and IT, integrated with business applications, with governance and monitoring. Like Microsoft, IBM supports both prebuilt and custom agents and multi-agent orchestration. For students, reading both companies' definitions is useful because enterprise buyers often use their terminology in requirements documents and job descriptions.
Are AI agents autonomous?
Some are, to a degree. Agents range from assistants that act only when a user asks, through semi-autonomous agents that propose actions for approval, to autonomous agents that run on triggers and complete tasks without a person starting them.
Also asked as: when ai agents started
Autonomy is a design choice, not a property of the model. Well-designed systems match autonomy to risk: an agent can freely search and summarise, but must ask before sending emails to customers, changing records or spending money. Autonomous agents still operate within the permissions, tools and limits their builders give them. In practice, most business agents today are semi-autonomous, with humans approving consequential steps and reviewing logs.
What happens when AI agents go rogue?
Agents "go rogue" when they take actions nobody intended: deleting data, sending wrong messages, looping and running up costs, or following malicious instructions hidden in content they read. The causes are usually design flaws rather than intent.
Common causes and defences:
- Excessive permissions: give each agent only the access it needs.
- Prompt injection: treat retrieved web pages, emails and files as data, never as instructions.
- Vague goals: write clear instructions and success criteria.
- No limits: cap steps, time and spending per run.
- No oversight: require approval for irreversible actions and keep full logs.
Test agents with adversarial inputs before launch; the OWASP list of LLM application risks is a good checklist.
Can AI agents be trained?
Agents improve mainly through better instructions, tools, knowledge and evaluation rather than retraining. The underlying model can also be fine-tuned for a specific domain or task, but that is less common than improving the system around it.
The improvement loop: review conversation logs and failures, fix instructions or tool descriptions, add missing knowledge, adjust which actions need approval, and re-run an evaluation set of tasks. Fine-tuning helps when the model must follow a very specific format or domain language consistently, or when a smaller, cheaper model must match a larger one on one job. Agents do not learn automatically from each conversation, by design, so their behaviour stays predictable.
Do AI agents have memory?
Yes, in two ways. Short-term memory is the conversation and working context within a task. Long-term memory is information stored deliberately, such as user preferences, past task summaries or case history, in a database or vector store and retrieved when relevant.
Long-term memory must be designed: what to store, for how long, who can see it, and how to correct or delete it. Storing too much clutters context and raises privacy risks; storing too little makes agents repeat questions. Platforms such as Copilot Studio offer variables and connections to storage; developer frameworks provide memory components. Treat memory like any personal data: minimise it, secure it and follow your organisation's retention rules.
What do AI agents run on?
AI agents run on three layers: a language model, usually hosted in the cloud or run locally; an agent runtime or harness that executes the loop, tools and limits; and the tools and data sources they connect to, such as APIs, databases and business applications.
In Microsoft's case, agents run in Microsoft's cloud within your tenant and Power Platform environments; with IBM, on watsonx; developers can run their own agents on any server or serverless platform with frameworks such as LangGraph or provider agent SDKs. Local agents can run on a laptop with open models through Ollama. Choose where to run based on data location, latency, cost and the systems the agent must reach.
Are AI agents SaaS?
Many are offered as SaaS: you subscribe to a hosted agent, such as a support agent or coding agent, and use it through a web app or integration. But agents can also be built in-house, embedded in existing products, or run on your own infrastructure.
Business models include per-seat subscriptions, usage-based pricing per conversation or action, and outcome-based pricing, such as per resolved ticket. Platforms like Copilot Studio meter usage in credits. Whether to buy SaaS or build depends on how specialised your process is, how sensitive the data is, and whether off-the-shelf agents fit. Many organisations start with SaaS agents for common tasks and build custom ones for processes that give them an edge.
Where can I sell AI agents?
Builders sell agents through cloud marketplaces such as Microsoft's and AWS's, through app stores for platforms like Salesforce and ChatGPT's app directory, as standalone SaaS products, or as custom builds for clients.
Also asked as: where to sell ai agents
Practical routes for students and small teams:
- Freelance and client work: build agents for local businesses, colleges or startups.
- Platform marketplaces: publish to the store of a platform your customers already use.
- Niche SaaS: one agent that solves one problem very well for a specific industry.
- Templates: sell or share workflows for tools like n8n.
Buyers care about reliability, data security and clear results, so include evaluation numbers, logging and a human fallback in what you sell.
When did AI agents start?
The idea of intelligent agents dates back decades in AI research, but agents built on large language models started around 2022 and 2023, with research such as ReAct and early open-source projects that chained model calls with tools.
Native tool calling in commercial models made agents far more reliable from 2023 onwards, and by 2025 coding agents, research agents and enterprise agent platforms from Microsoft, IBM, Google, Salesforce and others were mainstream. Standards such as the Model Context Protocol, introduced in late 2024, made connecting agents to tools easier. The term "agentic AI" became widespread in 2024 as the focus shifted from chatbots to systems that act.
Sources
- Microsoft Learn, Microsoft Copilot Studio documentationlearn.microsoft.com
- IBM, What are AI agents?ibm.com
- OWASP, Top 10 for Large Language Model Applicationsowasp.org
- Yao et al., ReAct: Synergizing Reasoning and Acting in Language Models (2022)arxiv.org
- Model Context Protocol, documentationmodelcontextprotocol.io
GenAI engineer and AI product builder with 2+ years shipping production-grade AI systems: RAG pipelines, fine-tuned LLMs, hybrid retrieval and multi-modal apps across vision, speech and OCR, architected end to end from ingestion to deployment. Leads engineering for SCULT INDIA's 14-member team, founded the 500+ member TechVerse Enclave community and has mentored 200+ students. Three hackathon first prizes: Changethon 2025 (IIT Roorkee), Product Genesis at Vividhotsava 2025 (CSJMU) and BYTEBATTLE (MeetKats).
Open to GenAI roles, hackathon judging, mentorship sessions and guest talks at colleges. For campus sessions the inviting institution covers the visit end to end: travel (a First AC train ticket for journeys up to five hours, a flight for anything longer), accommodation, meals, and pickup and drop; the honorarium and workshop fees are negotiable on WhatsApp or on the call, and requests are welcome from institutions that can cover the visit this way. Email pranjulrathour41@gmail.com.
