WEBVTT NOTE Sentence-level timing from the original narration; no invented word timing. 1 00:00:00.350 --> 00:00:07.070 When an AI reads a file, uses a tool, and checks a result, the model is only part of what you're seeing. 2 00:00:07.250 --> 00:00:10.050 The surrounding software makes that work possible. 3 00:00:10.230 --> 00:00:13.030 That surrounding system is called a harness. 4 00:00:13.210 --> 00:00:18.090 I'm going to show you one real task, and point out the pieces that make an agent useful. 5 00:00:18.890 --> 00:00:24.650 I'm using DeepSeek Harness, a real agent app with a conversation and tools for files. 6 00:00:24.830 --> 00:00:28.030 The model is one component inside this application. 7 00:00:28.210 --> 00:00:34.130 The instructions, available tools, and permission settings all affect what the complete system can do. 8 00:00:34.310 --> 00:00:39.670 The name in the model selector does not tell you everything about what the complete system can accomplish. 9 00:00:39.850 --> 00:00:43.130 The model chooses what to say or which action to request. 10 00:00:43.310 --> 00:00:46.910 A tool does something outside the model, such as reading a file. 11 00:00:47.090 --> 00:00:48.770 The harness connects those pieces. 12 00:00:48.950 --> 00:00:55.270 It supplies context, makes allowed tools available, runs permitted actions, and returns their results. 13 00:00:55.450 --> 00:00:59.450 The model can then use that new information to decide what to do next. 14 00:00:59.630 --> 00:01:02.990 For this task, I have a small file of fictional expenses. 15 00:01:03.170 --> 00:01:10.050 The goal is to calculate spending by category, flag anything that needs attention, and create a short summary. 16 00:01:10.230 --> 00:01:13.110 First, look at the information we are giving the agent. 17 00:01:13.290 --> 00:01:15.690 The file contains the original records. 18 00:01:18.247 --> 00:01:25.447 The instructions say how to handle a blank category: keep the record, group it separately, and do not guess. 19 00:01:28.004 --> 00:01:31.684 We also tell it to preserve the input instead of quietly changing the source. 20 00:01:31.864 --> 00:01:36.664 These details are context: the information available to the model for this job. 21 00:01:36.844 --> 00:01:39.484 More context is not automatically better. 22 00:01:39.664 --> 00:01:43.504 Unrelated files and old conversations can distract from the task. 23 00:01:43.684 --> 00:01:50.004 A useful harness helps select the relevant instructions and evidence, within the scope we allow. 24 00:01:50.184 --> 00:01:54.664 For longer tasks, a harness also has to manage limited context space. 25 00:01:54.844 --> 00:01:58.524 It may select earlier results or summarize older conversation. 26 00:01:58.704 --> 00:02:03.184 Summaries can lose detail, so keep critical facts in an artifact you can reopen. 27 00:02:03.364 --> 00:02:09.204 The file is durable; whether the next response uses it depends on what the system actually loads. 28 00:02:09.384 --> 00:02:16.264 Before pressing send, check that the request names the input, the expected output, and the rule for uncertainty. 29 00:02:16.444 --> 00:02:19.244 That makes success something we can actually inspect. 30 00:02:19.784 --> 00:02:21.224 Now we submit the task. 31 00:02:21.404 --> 00:02:24.764 Watch the activity between the prompt and the final response. 32 00:02:24.944 --> 00:02:28.784 The agent requests a tool, and the application performs the action. 33 00:02:28.964 --> 00:02:34.084 Opening the tool details lets us see which file it read and what information came back. 34 00:02:36.642 --> 00:02:38.402 We can see what it actually read. 35 00:02:38.582 --> 00:02:41.302 The model uses that result to choose its next step. 36 00:02:41.482 --> 00:02:46.042 If it needs another file or a calculation, another tool request can follow. 37 00:02:46.222 --> 00:02:51.982 This is the basic agent loop: decide, act through a tool, inspect what happened, and continue. 38 00:02:52.162 --> 00:02:54.882 This example uses files and a calculation. 39 00:02:55.062 --> 00:03:01.862 Those capabilities come from the integration around the model, and they should be limited to what the task needs. 40 00:03:02.042 --> 00:03:06.522 Installing a connector does not mean the agent should have unrestricted access. 41 00:03:06.702 --> 00:03:09.822 The loop does not guarantee that every decision is correct. 42 00:03:10.002 --> 00:03:14.962 It gives the system a way to work with evidence instead of stopping at a single response. 43 00:03:15.142 --> 00:03:20.822 In this example, the useful evidence is the original expense data and the calculations made from it. 44 00:03:21.002 --> 00:03:25.722 We can follow the activity and compare the proposed summary with those records. 45 00:03:25.902 --> 00:03:31.742 Notice that an empty category is still something to flag, not an invitation to invent a label. 46 00:03:31.922 --> 00:03:35.202 Check whether the actions and results support the answer. 47 00:03:35.382 --> 00:03:38.902 Before accepting the summary, we check it against the original data. 48 00:03:39.082 --> 00:03:45.962 First, the calculator totals the source records. At this stage, it correctly reports that no summary has been saved. 49 00:03:46.142 --> 00:03:49.982 The expense with a missing category must remain visible in that check. 50 00:03:50.162 --> 00:03:55.442 Twelve records total two hundred and twenty-seven, including seventeen without a category. 51 00:03:57.999 --> 00:04:02.959 The calculated totals match the source records; that is evidence for this calculation. 52 00:04:03.139 --> 00:04:04.979 Now look at the write boundary. 53 00:04:05.159 --> 00:04:09.079 Reading the input and saving a new file are different actions. 54 00:04:09.259 --> 00:04:15.259 The first write is blocked by the read-only setting. The agent then asks for one-time permission to save this summary. 55 00:04:15.439 --> 00:04:22.719 Here, the application restricts which files and command the agent can use, and asks before writing this summary. 56 00:04:22.899 --> 00:04:26.099 This demonstrates an application permission boundary. 57 00:04:26.279 --> 00:04:31.399 An operating-system sandbox is a separate boundary, and we have not demonstrated one here. 58 00:04:31.579 --> 00:04:35.979 Inspect the proposed destination and contents before allowing this specific action. 59 00:04:38.536 --> 00:04:43.096 After the file is saved, reopen it and confirm what actually exists. 60 00:04:43.276 --> 00:04:50.476 The checker compares the saved category totals, grand total, and the count of records missing a category against the source. 61 00:04:50.656 --> 00:04:55.536 A pass supports those checks; it does not verify every sentence in the report. 62 00:04:55.716 --> 00:04:59.556 That final readback matters more than a message saying the work is done. 63 00:05:02.113 --> 00:05:06.433 Saved files only become context when the system loads them again. 64 00:05:10.113 --> 00:05:13.473 Approvals and sandboxes also do different jobs. 65 00:05:13.653 --> 00:05:16.693 An approval asks whether an action may proceed. 66 00:05:16.873 --> 00:05:20.393 A sandbox limits what the execution environment can reach. 67 00:05:20.573 --> 00:05:22.893 Always check what is actually enforced. 68 00:05:23.073 --> 00:05:29.393 The original file is unchanged, the totals have been checked, and the saved summary can be reopened. 69 00:05:31.950 --> 00:05:35.950 Following those steps gives us a practical way to evaluate the system. 70 00:05:36.130 --> 00:05:39.490 When you evaluate an agent, ask four questions. 71 00:05:39.670 --> 00:05:41.430 What information can it see? 72 00:05:41.610 --> 00:05:43.130 Which actions can it take? 73 00:05:43.310 --> 00:05:44.590 How is the result checked? 74 00:05:44.770 --> 00:05:46.210 And where must it stop? 75 00:05:46.390 --> 00:05:50.230 Those questions tell you much more than the name of the model alone. 76 00:05:50.410 --> 00:05:52.890 Try this with one small task of your own. 77 00:05:53.070 --> 00:05:59.550 Follow the actual steps, inspect the output, and identify one check you would want before trusting the result.