Lesson 3 of 5
Who decides whether an action runs?
The trace ended with the harness waiting. The model asked to email a supplier, and nothing in the request mentioned email. Whether that message goes depends on three different things: what the model chooses to propose, what the harness allows, and what software outside both makes impossible. This lesson separates the three, then asks what a person's approval actually adds.
Numbered links identify sources. Underlined terms offer a definition without leaving the lesson.
Three layers: the model's choice, the harness's check, the enforced boundary
Module 2 ended its access lesson with a distinction this module needs from the first paragraph: an instruction in the conversation asks the model to behave, while an is a limit that software other than the model applies. Agents add a layer between those two, so there are three places where an action can be prevented.
The first layer is the model's choice. The request, the application's own instructions and any brief you write shape what the model proposes. Module 1 explained why following an instruction is likely and not guaranteed. On line 19 of the trace, nothing had asked for an email; the model proposed one as a way to finish the task. An instruction such as “do not contact anyone” would have made that proposal less likely. It would not have made it impossible.
The second layer is the harness's : before it runs a tool call, the harness consults rules that say, for that tool or that kind of input, whether to run it, ask a person first, or refuse. This is software, so it applies whether or not the model “meant” the call, and it applies only to what its rules name. A rule that asks before every email does nothing about a file write. A harness can also leave a tool out entirely, so the model is not offered a callable operation by that name. Those are three different states for one capability: unavailable, asked about, or discouraged by an instruction, and only the first two are enforced.
The third layer is enforcement outside the harness: the authorization of the connected account, the operating system's file permissions, a network rule. One common form has a name you will meet in product descriptions. A is an enforced restriction that confines a process to specified files, network destinations and other resources, applied by the operating system or by running the process inside a container or a virtual machine, so that an operation outside the boundary fails even if the harness tries it. Boundaries need a stated scope and a correct implementation. A sandbox that blocks file writes may still permit a network request; inspect each route through which the agent can act.
Apply the three layers to line 19. Nothing in the request asked for email, and the model proposed it anyway. The harness's rule for send_email was to ask, which is why line 20 waits; had the rule been to allow, the email would already be gone. The connected account is authorized to send, and nothing outside blocks it. So the only thing between the proposal and a sent email is the organizer's answer. The same analysis covers a second kind of out-of-task proposal. Suppose that at line 16 the model had also proposed rewriting fencing-brief.txt to correct a typo it noticed. That edit is outside the task. Under this example's rules it would produce a prompt; under a harness set to accept edits inside the working folder without asking, it would simply run.
One product's controls, dated
A dated example makes the layers concrete. The documentation for Claude Code, read on 29 September 2026, describes a permission system with rules of three kinds: allow, ask and deny. In the mode it labels Manual, the harness asks before most actions that edit files, run shell commands or reach the network, while file reads inside the working folder run without asking. 4 5 The same page states the layer distinction in its own words: permission rules are enforced by Claude Code and not by the model, and instructions in a prompt or an instruction file shape what the model tries to do without changing what the harness allows. A deny rule that names a whole tool removes it from what the model is told about, which is the “unavailable” state from the previous section. 4
The documentation also lists other modes on that date. One accepts file edits in the working folder without asking. Plan mode is intended for reading and exploration, but current documentation also permits classifier-approved commands in some configurations. Auto mode uses a separate model to review actions routed to it, while explicit ask rules still prompt the person. One denies anything that would otherwise ask. And one bypasses the prompts, which the documentation says to use only in isolated environments where the harness cannot damage the host. 5 This is one product's set of controls on one date, and it will change. What transfers is the shape: named rules, modes that change which rules ask, and a boundary that the product's own documentation says lives outside the model. When you meet a different harness, look for those three things and read what each covers.
What an approval prompt shows, and what it cannot
A useful approval prompt shows the proposed operation before it runs: the tool, the address, subject and body of an email; the path and contents of a file write. That is the moment module 2 described as checking before acting, and it is an opportunity to catch mistakes before an external effect occurs. Reading the prompt is how you compare the proposed operation with what you meant.
Even a detailed prompt leaves questions unanswered. Interfaces also differ in how much of the operation they reveal. It cannot show whether the operation will achieve the task; the write at line 17 was approved, and the table could still be wrong. It cannot show what the model will propose next. And it cannot show consequences after execution, such as a recipient keeping a copy. Its value also depends entirely on the person reading it.
Anthropic's May 2026 account reports that Claude Code users approved about 93 percent of permission prompts. The company describes approval fatigue as a reason to combine oversight with enforced boundaries. It also reports a boundary failure: an allowed network destination could be used to send data out for an unintended purpose. These are vendor observations, not a test conducted for this course. 3
Compare two authored designs. One asks before every email; another has no sending capability. The first lets a person authorize a useful message but depends on their decision. The second prevents this task from sending messages, at the cost of making that operation unavailable. Neither tells you whether a written comparison is correct. Choose controls for the particular effects that matter, and check the resulting work separately.
A role is not a permission
Agent products and practitioner reports often describe agents by role: a reviewer, a researcher, a read-only assistant. A role stated in instructions is a first-layer fact. It grants nothing and removes nothing. Whether a reviewing agent can change files depends on whether a write tool exists in its harness, whether the rules allow it, and whether anything outside the harness blocks it. The same applies to the sentence “the agent only has access to this folder.” Ask which layer that describes: an instruction, a rule in the harness, or a boundary the operating system or a container enforces.
This is not a reason to distrust roles. Instructions are how you shape what a model proposes, and module 4 is about writing them well. It is a reason to keep the layers separate when someone tells you an agent is safe because of what it was told.
Oversight is a decision about where to look
Put the layers together and human oversight stops meaning watching every step. It means deciding, before the loop starts, which actions may run freely, which must ask, and which are denied or made unavailable; reading the prompts that do appear; and reading the action record and the output afterwards. For the fencing task that could be: reads run freely, the one file write asks, and email is not offered at all. With those settings the organizer reads one prompt, at line 17, and any request for that unavailable tool is rejected. The next lesson turns this into a written brief that names those settings alongside the task.
Carry this forward
- Three layers decide whether an action runs: what the model proposes, shaped by instructions; what the harness's rules allow, ask about or deny; and what software outside both enforces.
- A permission check is software: it applies regardless of the model's intent, and only to what its rules name. Leaving a tool out removes that callable operation; asking means a person decides each time; an instruction only makes a proposal less likely.
- An approval prompt shows the operation as it will run and nothing about its outcome. Its value depends on the person reading it; one vendor's telemetry reports most prompts approved and attention falling as prompts multiply.
- A role in instructions grants and removes nothing. Ask which tools exist, which rules apply and which boundary is enforced, and what each covers.
Lesson check
Choose an answer for each question.
This is a practice check. Answers reset when you leave this page. Results are not saved and do not place you in another module or award a qualification.
You now know what to look for in a harness's controls. The next lesson turns that into a decision: which work to hand over, and how to write the handover so the controls and the checks are set before the loop starts.
Sources · Reviewed 2026-09-29
- 3. Anthropic: How we contain Claude across products
- 4. Claude Code documentation: Configure permissions
- 5. Claude Code documentation: Choose a permission mode
Sources support nearby explanations. Product details can change. Scenarios and model results are authored illustrations unless a lesson explicitly identifies a recorded execution. Module 7 identifies its offline code checks.