Free tools Windows power users keep installed
One-click scans. No signup required.
A coding agent is more than a language model generating code. The model proposes what to do next; an agent harness supplies context and tools, runs permitted actions, feeds the results back to the model, and keeps track of the work. That repeated exchange is how an agent can inspect a repository, respond to errors, and leave changes in a workspace—not just produce a block of text.
The useful distinction is between the model’s decisions and the software that makes those decisions actionable. Once you see that division, it becomes easier to understand why two agents using similar models can behave differently: their tools, context, permissions, execution environments, and runtime design may not be the same.
As an Amazon Associate I earn from qualifying purchases.
How does a coding agent actually work?
A coding agent typically runs a loop: the harness sends the model a request with instructions and relevant context; the model returns either a response for the user or a request to use a tool. If there is a tool request, the harness checks and executes it, then supplies the result to the model for another step.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errors- Prepare: The harness combines the user’s request with applicable instructions, conversation history, and available context.
- Ask the model: The model returns a user-facing answer or an action request, such as inspecting a file or running a command.
- Execute: The harness routes an allowed action to the appropriate tool or environment. Depending on the system’s rules, an action may require approval or may be unavailable.
- Return the result: The tool’s output is added to the ongoing interaction so the model can interpret it and decide what to do next.
- Continue or finish: The cycle repeats until the model provides a response rather than requesting another action.
For example, a request to fix a failing test might lead the model to ask for the test output, inspect the relevant code, propose an edit, and run the test again. The failure message from one step can affect the next. The final result may therefore include both an explanation and modifications to files in the workspace. OpenAI describes this repeated interaction as the agent loop.
#1 Best Overall
The model is not directly reaching into the computer by itself. It emits an action request in a format the harness can handle; the harness decides how that request is routed and what result comes back. The exact mechanics vary by system, but the division between proposing an action and executing it is fundamental.
What is an agent harness?
An agent harness is the software layer that wraps a model and makes it work as an agent. Microsoft’s explanation describes the model as making reasoning and action-request decisions while the harness turns them into a stateful workflow and tracks the conversation and changes.
In practice, the harness may prepare the model’s request, expose tools, route tool calls, return their results, apply permission and approval rules, and maintain session state. It determines much of the agent’s operating environment: what information is available, which actions are possible, and how the system handles progress between steps.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →A research framework for harness responsibilities
A July 2026 source-code study grouped observed harness responsibilities into seven areas: the agent loop, model integration, tools and actions, memory and context, safety and permissions, orchestration, and extensibility. The authors examined eleven selected systems; that sample is a framework for understanding those systems, not a census of every coding agent or a universal industry standard.
The study also distinguishes an agent harness, which wraps a model to enable action, from an evaluation harness, which wraps an agent to run it against tasks. Similar terminology does not mean they serve the same purpose.
What happens when an agent uses a tool?
A tool is an action surface exposed to the model. It might let the agent read or edit files, run a shell command, or access a service. The model can request a tool when its description and schema make that option available; the harness or service then handles the request and returns a result.
Anthropic’s tool-use documentation describes a common pattern: define a tool’s schema, implement a handler or callback, return its result, and allow the model to decide when the function is appropriate. A schema specifies what the tool accepts; the handler determines what actually happens. Some tools are executed by an external service rather than by the application hosting the model. In that design, the service may perform several internal steps before returning a result, and an iteration cap can pause the work and require continuation.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Tools do not have to appear to users as separate buttons. They can be internal capabilities managed by the harness, and the details of the interface are a design choice. An empirical study of coding-agent harness design reports that predefined tools can help models with weaker bash proficiency, while models able to use bash can work effectively with a bash-only interface—and at lower cost on command-line-centric tasks in the evaluated setup. That finding is specific to the study’s conditions; it does not establish one tool design as best for all models or tasks.
How do context, state, and workspace affect the work?
Context is finite
A model’s context window has a limit and includes both input and output tokens. As a task continues, conversation history and tool results can accumulate. The runtime therefore has to decide what to keep available, what to summarize, and what information to bring forward. If relevant output is no longer present in the context, the model may not be able to use it in a later step unless the harness retrieves or reintroduces it.
State preserves progress
State is the information a system keeps about an ongoing or previous task: for example, conversation history, session configuration, or details needed to resume work. How state is stored and reused depends on the runtime. It may be managed by a service, maintained by an application, or assembled by the developer when making direct model calls.
A workspace gives the agent somewhere to act
A prompt can be enough when a task is limited to reasoning over supplied text. Work that depends on files, commands, packages, or artifacts usually needs an execution environment with a workspace. OpenAI’s sandbox guide describes capabilities such as files, commands, packages, mounted storage, exposed ports, snapshots, and resumable state; a particular sandbox need not provide all of them.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
A helpful architectural distinction is the control plane versus compute. The harness can coordinate model calls, tools, approvals, tracing, recovery, and run state, while a separate sandbox executes model-directed work against a filesystem and command environment. Keeping the two apart can let trusted infrastructure retain credentials, billing, audit records, review, and recovery responsibilities while execution happens in an isolated environment.
Why does a coding agent need permissions and a sandbox?
Permissions determine which actions may run, which need approval, and which are blocked. They are a harness responsibility, not an automatic property of the model. The execution environment is a separate choice: it can be provider-managed, self-hosted, or unnecessary for tasks that do not involve persistent workspace operations.
A sandbox can isolate file and command execution, but the word alone does not establish that a system is safe. The design still needs to specify what files and services the environment can reach, which credentials it receives, how actions are approved, and how changes are reviewed. A useful boundary is to give the execution environment only the access it needs, while keeping sensitive orchestration functions and credentials in trusted infrastructure where possible.
For engineering teams, the practical aim is not to remove all risk but to make the boundaries explicit: define permitted actions, make consequential operations reviewable, and ensure changes can be inspected. These are design recommendations, not guarantees supplied by the presence of a sandbox.
How do common agent runtime approaches differ?
OpenAI’s Agents documentation distinguishes three approaches by how much of the runtime the provider manages versus the application. These are examples of architectural choices, not rankings; the suitable balance depends on required control, state handling, and workspace needs.
Best Value
| Approach | Who owns orchestration? | State between tasks | Tool execution and environment | Typical fit |
|---|---|---|---|---|
| Agents API | Managed Codex harness; OpenAI manages state and infrastructure. | Managed by the service for longer-running work. | Uses the managed runtime and its available execution capabilities. | Teams that want a managed harness for longer-running work. |
| Agents SDK | The application controls deployment, storage, approvals, and runtime integration; the runner handles the loop and handoffs. | Application-controlled. | Integrated through the application’s chosen runtime and tools. | Teams that want to control deployment and connect the agent to their own systems. |
| Responses API used directly | The application builds more of the integration around direct model calls. | The application manages history and chaining. | Can combine available hosted tools with the application’s own execution environment. | Teams that need direct control over how the model interaction and surrounding workflow are assembled. |
The table describes the distinctions in OpenAI’s documentation; the available tools and execution details depend on the implementation. In particular, selecting an API or SDK does not by itself determine whether a task has an isolated, persistent workspace. That is a separate runtime decision.
Questions to ask before choosing
- Orchestration: Do you want a provider-managed workflow, an application-controlled runner, or direct model calls with more integration work?
- State and resume: Who saves the session information, and how will a long task continue after interruption?
- Tools and compute: Which actions are hosted, implemented as application callbacks, or executed in your own environment?
- Workspace: Does the task need files, a shell, packages, persistent artifacts, or resumable work?
- Permissions and review: Where do approvals, credentials, audit records, and execution isolation live?
What makes a coding-agent workflow easier to trust?
OpenAI’s account of its agent-first engineering workflow describes using repository tools and embedded skills to gather context, reviewing changes locally, requesting additional targeted reviews, responding to feedback, and iterating. It also argues for enforcing architectural invariants while leaving implementation choices open. Those are practices from OpenAI’s own workflow, not independently established rules for every team.
As an engineering judgment, a useful workflow makes the repository context relevant to the task, keeps the action surface appropriately scoped, preserves enough state to continue coherently, and makes risky changes subject to permission or review. The finished work should also be checkable—for example, by inspecting the files changed and the results of relevant commands. A harness can coordinate these steps, but the team still has to decide what counts as acceptable evidence of a correct change.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
