Your AI agent needs a receipt

A beginner’s guide to the record that shows what an AI agent saw, did, and returned.

Before we talk about logs, what is an AI agent? An AI model usually gives you an answer. You ask a question, it responds, and the conversation ends there. An AI agent goes a step further : it can work through a task, use tools, look up information, and take an action. Think about a support inbox. A normal chatbot might suggest a reply. An agent could read the message, search the refund policy, check an order, draft the reply, and decide whether a human needs to review it. The model is the brain that makes decisions. The agent is the larger system that connects that brain to tools and real work. That extra ability is what makes agents useful. It is also what makes their mistakes harder to understand. When a chatbot gives a weak answer, you can try again. When an agent changes a record, sends the wrong reply, or uses the wrong document, you need to know exactly how it reached that point. A log is the agent's receipt A log is simply a written record of what happened. Your bank statement is a log of transactions. A delivery tracker is a log of where a package travelled. The history in a shared document is a log of who changed what. An agent log should do the same thing for AI work. It should record what the agent received, which information it used, which tools it called, what those tools returned, what answer it produced, and what action happened next. If AI can take an action, it should leave a receipt. Without that receipt, a successful demo can be misleading. You see the final answer, but you cannot tell whether the agent found the correct policy, used an old file, retried a broken tool five times, or simply guessed. The output looks clean while the process underneath it may be fragile. What changed this week On August 4, Simon Willison released LLM 0.32. Despite the name, LLM is not a new AI model. It is an open-source tool that developers use to run different models from the command line or from Python. The new version changes how it records conversations, tool activity, and model responses. The release introduces a content-addressed message store in SQLite. In plain English, SQLite is a small database that can live in a single file. Content-addressed means a message is given a digital fingerprint based on its contents and its place in the conversation. When the same conversation history appears again, the system can point to the existing messages instead of storing the same material repeatedly. Why should a non-developer care? Better records make it easier to compare two runs of an agent. You can ask whether the input changed, whether a different policy was retrieved, whether the model changed, or whether the tool returned something new. That is much more useful than staring at two different final answers and guessing what happened. LLM 0.32 also adds visible reasoning summaries for models that support them, more support for tools run by AI providers, and new OpenAI Responses API features. Those are useful technical improvements. The beginner-level idea is simpler: the tool now keeps a clearer record of the work around an AI response. An example you can picture Imagine an online store gives an agent permission to help with refund requests. A customer says an order arrived late and asks for their money back. The agent reads the message, finds the refund policy, checks the order details, and recommends a response. Now imagine the recommendation is wrong. Saying “the AI made a mistake” does not help the support manager. They need to know which version of the policy the agent read. Did it find the correct order? What date did the order tool return? Did the agent ask the tool the right question? Was the reply only drafted, or was it actually sent? A useful log answers those questions. It may reveal that the model behaved reasonably but received an outdated policy. It may reveal that the order tool returned the wrong timezone. It may reveal that the agent skipped a required approval. These are different problems, and each one needs a different fix. This is why the interface is not the most important part at the beginning. A beautiful dashboard can show a green checkmark and a confident explanation. If the underlying record is incomplete, the dashboard is only decorating uncertainty. What a useful agent log should show You do not need to understand database design to review an agent. Start with five ordinary questions: What did it receive? The request, relevant customer or task details, and the version of the instructions. What information did it use? The documents, search results, or records it retrieved. What did it do? Every tool call, the information sent to that tool, and the result returned. What came out? The draft, decision, classification, or action produced by the agent. What did the run cost? The model used, time taken, token usage, errors, and retries. Use this as a quick check for any AI workflow. Reasoning summaries can provide extra clues, but they are not the final proof of what happened. The strongest evidence is the action record: which tool was called, what data went into it, what came back, and what the system did afterward. The simple rule LLM 0.32 does not automatically make every agent safe or reliable. It gives builders better raw material for understanding their systems. The same principle applies even if you never use Simon's tool: save enough information to explain the agent's work later. The practical order is straightforward. First record the inputs, sources, tool activity, outputs, and final action. Then create a few simple ways to search those records. Build the polished dashboard after you know the evidence underneath it is complete. Agent demos show what happened when everything went well. Agent logs help you understand the day something goes wrong. Before giving an agent more freedom, make sure it can leave a clear trail behind it. If one of your AI workflows made a bad decision tomorrow, could you explain exactly what it saw and did?

Harshith Vaddiparthy works with founders, operators, and teams on practical AI products, workflows, advisory, training, and mentorship. This no-JavaScript version preserves the page's core information and navigation.