Why are AI agents so slow, and what speeds them up?
AI agents are slow because they do not answer in one go. An AI agent is a program that uses AI to carry out a task for you, and it works in rounds: it reads, decides, looks something up and starts again. Every round takes a few seconds. Most of the waiting goes away when the agent has less to read and a clearer task.
A comparison from the office
Imagine a temp on their first day. You ask: "Which customers have we not contacted for two weeks? Write each of them a short note." The temp walks to the filing cabinet, comes back, reads their notes from the top, walks to the cabinet again and so on. Opening the cabinet is quick. The walking, rereading and writing take the time.
An agent works the same way. Looking something up in your data is fast. Reading and writing are slow, and the agent does both again in every round.
Where the time goes: one request, step by step
Take the same request about customers without contact. This is what happens before you see an answer.
- The agent reads its instructions and your question.
- It decides to look up the open deals and asks your customer system for them.
- The list comes back. This is usually the quickest part.
- The agent reads the list and sees that it needs the last contact for each deal. It looks those up, often one after another.
- It reads everything again: instructions, question and all results so far.
- It writes the notes, word by word.
Your data was asked a few quick questions. The AI was called four or five times, and each time it had more to read. That is the pattern behind almost every slow agent: many rounds, and a lot of text in each.
Four common causes
- A vague task. "Find the deal I mean" makes the agent search around. Every extra look costs a full round.
- Too much to read. Whole documents and long histories are read again in every round.
- A long answer. Writing is the slowest thing an AI does. A page of text takes far longer than a yes or no with one line of reasons.
- A stronger AI than the task needs. The most capable AI models take longer to answer. A simple check rarely needs one.
What speeds an agent up
- Say exactly where to start. "This deal" is faster than "one of our deals".
- Give it the passage, not the file. Three paragraphs that matter are read faster than thirty pages.
- Ask for a short answer. A category, a date and one sentence are written quickly.
- Use a fast AI model for simple jobs. Keep the strong one for tasks with many steps.
- Take fixed steps away from the agent. Anything that is always the same belongs in a simple rule, which runs at once.
When slow does not matter
Speed matters when a person sits in front of a chat and waits. It matters much less when the agent works in the background. If a new enquiry is sorted one minute after it arrived, nobody notices the minute. So first ask who is waiting. Long jobs, such as going through a whole customer list, can run at night or start by themselves when something changes. The result is then simply there when people need it.
How Worknano handles this
In Worknano, agents and the assistant use semantic search, which finds text by its meaning. They read only the passages that matter, so answers arrive faster. Each agent has its own model setting: a fast model suits short checks, a balanced one suits tasks with several steps. Agents can also run by themselves when a record changes or on a schedule, so longer jobs keep nobody waiting. The audit log records how long every agent run took, which shows you which one to improve.
- Semantic search across all your work
- AI agents in Worknano
- Build your own agent: model, tools and instructions
Create a free workspace and time an agent on your own records.