Tools, Memory and Protocols: What Changed in Agentic AI This Year

· AI Technology · 7 min read

Agents stopped being chat windows and became software that uses your tools. A plain-English tour of tool use, standard connection protocols, long-term memory and computer-use agents.

The shift: from talking to doing

The biggest advance in agentic AI is not that models got smarter at writing. It is that agents can now reliably act: read your calendar, write to your CRM, pull an invoice, send an email, update a spreadsheet. The conversation is now the interface to your systems, not a separate box on the side.

Tool use, made dependable

Modern agents are given a list of actions they may take, with strict definitions of what each one needs. Instead of hallucinating an answer about stock levels, the agent calls the stock lookup and reports what came back. The practical effect for a business is that answers become verifiable — every claim traces to a system, not to the model's imagination.

Standard connection protocols

Until recently, every integration was bespoke. Open standards for connecting models to tools and data have changed that: one connector to your calendar, one to your CRM, reusable across agents and vendors. It matters commercially because it lowers switching costs — you are connecting your systems once, not marrying a single AI vendor.

Memory that survives the conversation

Agents now keep durable memory: who this customer is, what they asked in March, which quote they rejected and why. Two rules keep it useful and safe:

  • Store facts, not transcripts. "Prefers WhatsApp, budget around SGD 5k, needs delivery by Q4" beats 40 pages of chat.
  • Make it reviewable and deletable. If a customer asks what you hold, you should be able to show and remove it.

Computer-use agents

Some agents can now operate a browser or desktop directly — clicking through a portal that has no API. It is genuinely useful for legacy systems, and genuinely fragile: interfaces change, sessions expire, and a misclick has real consequences. Use it for read-heavy tasks and low-risk data entry, always with logging and a human approval step for anything that spends money or contacts a customer.

Longer horizons and planning

Agents can now hold a multi-step plan over hours or days: chase a lead on day one, follow up on day three, escalate on day seven, close the loop on day fourteen. This is where most SME revenue actually hides — not in the first reply, but in the fifth.

What to do with all this

  1. Pick the one workflow where speed or persistence loses you money.
  2. Give the agent read access first, write access second, send access last.
  3. Log every action, and review the log weekly for the first month.
  4. Keep a human approval gate on anything irreversible.

Advances only matter when they land in a workflow. The businesses getting value are not the ones with the newest model — they are the ones who connected an agent to the enquiries already sitting in their inbox.