Agent Claims Decoded
"We have agents" is not an architecture description. This eight-part series gives executives, architects, engineers and buyers a neutral way to test the claim. It sets out seven practical levels from chatbot to multi-agent system and the minimum closed loop that makes a system agentic, compares chatbots, copilots, AI workflows and bounded agents, explains how an agent chooses its next step and how it must stop, separates what an agent can do from what it is authorised to do, distinguishes run state and case threads from long-term memory, shows where a person must genuinely intervene, asks when several agents are better than one, and ends with a twelve-domain question bank and evidence scorecard. It follows Ubuntu Municipal Services, a fictional South African municipal services entity piloting a service-resolution agent for missed-waste-collection complaints; its systems and figures are illustrative. Research claims were checked on 3 November 2026.
8 articles
- Part 1Everyone says they have agents. What do they actually have?The word agent now covers everything from a chat window to a system that changes records on its own. Part 1 of Agent Claims Decoded gives executives, architects and buyers a neutral test: seven practical levels, the minimum closed loop, and the components a production agent cannot do without.10 min read
- Part 2A chatbot, a copilot, a workflow and an agent walk into the enterpriseThe same model can sit behind a chatbot, a copilot, an AI workflow and a bounded agent. What differs is who starts the work, who chooses the next step and whether records change. Part 2 of Agent Claims Decoded compares the four patterns on behaviour, authority, failure and evidence.7 min read
- Part 3The agent loop in plain English: how it chooses, and how it stopsAn agent's loop is easy to draw and hard to govern. Part 3 of Agent Claims Decoded explains, without jargon, how goal, action, observation and adaptation fit together, why a dynamic path is harder to test than a fixed one, and the five exits every loop needs.7 min read
- Part 4Agent capability is not agent authorityWhat an agent can technically do is set by its tools and accounts. What it may do is a decision the organisation must make, action by action. Part 4 of Agent Claims Decoded sets out a five-level authority model and shows why the gap between capability and authority is where agent incidents start.7 min read
- Part 5Memory is not stateWhen a supplier says its agent 'has memory', it usually means long-term recall across conversations. What an operational agent needs first is state: a checkpoint for the current run and a case thread for the current piece of work. Part 5 of Agent Claims Decoded separates the three and shows which failures each prevents.7 min read
- Part 6Human-in-the-loop is not a buttonAn approve button is not oversight. Meaningful human intervention needs the right point in the process, the evidence to decide, the authority to refuse and the time to look. Part 6 of Agent Claims Decoded gives a simple test for where a person must step in, action by action.7 min read
- Part 7Multi-agent is an organisational choice, not a feature countFive agents on a slide do not make a system more capable. Separate agents earn their place when they create a real boundary: different permissions, an independent checker or a different evaluator. Part 7 of Agent Claims Decoded gives a decision canvas for splitting, and the coordination costs every split adds.7 min read
- Part 8They say they have agents. Ask these questions.'We have agents' is the start of due diligence, not the end of the architecture discussion. Part 8 of Agent Claims Decoded is a reusable question bank across twelve domains, with the evidence to request for each and a four-level scorecard for rating the answers.11 min read
Working through any of this in your own organisation?
These articles come out of delivery work rather than theory. If the problems in this series look familiar, a short scoping conversation is usually the quickest way to work out what is worth doing first.
Talk to CloudNala