iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
A coding agent works through a loop: a model decides what to do and may request a tool; the harness runs that action, returns the result, and gives the model another chance to respond or act. The harness—not the model alone—connects reasoning to a workspace, manages context and permissions, and tracks the work as it happens.
How does a coding agent actually work?
OpenAI describes the repeating process as an “agent loop.” A simplified coding-agent loop looks like this:
- Prepare the request. The harness combines the user’s request with applicable instructions, conversation history, and available tool definitions.
- Ask the model. The model receives that context and returns either a user-facing response or a request to use a tool.
- Run the requested action. If the model requests a tool, the harness routes the request to the appropriate implementation, subject to the system’s permissions and approval rules.
- Return the result. The harness adds the tool’s result to the ongoing interaction and calls the model again.
- Continue or finish. The model may request another action, or produce a final response. The loop ends when it has no further tool action to request.
A tool result can change what the model does next. For example, a command might reveal the repository’s files or report an error; the model can use that information to choose a follow-up action. Tool calls can also change the workspace. The result of an agent run may therefore include both a final message and changes to files or other artifacts—not just text in the conversation.
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWhat is an agent harness?
An agent harness is the software around a model that makes it possible to use the model as an agent. Microsoft describes the division this way: the model makes reasoning and action-request decisions, while the harness turns those decisions into a stateful workflow and tracks conversation and changes. The model can request an action, but the harness determines how that request is handled in the application.
#1 Best Overall
In practice, the harness can prepare model inputs, expose tools, route tool calls, return results, apply permissions, and maintain session state. It may also coordinate approvals, recovery, and other parts of a run. The precise division depends on the runtime: “harness” describes a role in the system, not one universal product architecture.
A July 2026 source-code study of eleven selected coding-agent systems grouped observed harness responsibilities into seven areas: the agent loop; model integration; tools and actions; memory and context; safety and permissions; orchestration; and extensibility. That is one research framework based on a selected corpus, not a settled industry standard or a count of all coding agents. The study also distinguishes an agent harness, which enables a model to take actions, from an evaluation harness, which runs an agent against tasks.
What happens when an agent uses a tool?
A tool is an action surface made available to the model. It might let the agent inspect or edit files, run a shell command, use a browser, or call a service. Tools do not have to appear as separate buttons to the user: the harness can make them available behind the scenes, then decide how a request is executed.
Free tools Windows power users keep installed
One-click scans. No signup required.
A common pattern is to describe a tool with a typed schema, provide an implementation that handles the request, and return a result for the model to consider. Anthropic’s documentation describes this contract: define the schema, handle the callback, return the result, and let the model decide when the function is appropriate. In server-executed tools, a service may carry out multiple internal steps before returning. An iteration limit can pause that work and require it to continue later.
The available tools shape the agent’s practical capabilities. A model can only request actions the system makes available, and a tool request is not the same as permission to perform any conceivable action. The harness and execution environment determine what actually runs and under what conditions.
Choosing the action surface
Tool design involves trade-offs. Predefined tools can give a model a more structured way to act. An empirical study of harness design reports that predefined tools helped models with weaker bash proficiency, while bash-capable models worked effectively with a bash-only interface and lower cost on command-line-centric tasks in the study’s evaluated setup. That finding is specific to the models and tasks evaluated; it does not establish that one interface is best for every agent.
How do context, session state, and workspace differ?
Context is the information available to the model for a particular call. It can include instructions, conversation history, and tool results. A model’s context window is finite and includes both input and output tokens. As a task runs, conversation and tool output can accumulate, so the runtime has to decide what to keep available, summarize, or otherwise manage.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsSession state is the information the runtime preserves about the ongoing work, such as conversation and changes. Whether and how that state persists depends on the runtime design. A new model call does not, by itself, guarantee that an application has preserved everything needed to resume a task.
The workspace is where execution can inspect or change files and other resources. A sandbox may provide a filesystem, commands, packages, mounted storage, exposed ports, snapshots, or resumable state. These capabilities matter when the task requires operating on a repository, producing artifacts, or continuing work in an environment. For a short answer that can be produced from the prompt alone, a persistent workspace may not be necessary.
These are related but distinct concerns: context is what the model can use in a call, session state is what the runtime tracks across the work, and a workspace is where actions can operate. A system can preserve session state without making every detail fit in one model call, and a workspace does not automatically determine how much context the model receives.
Why does an agent need a sandbox?
A sandbox gives an agent a bounded environment for commands and workspace operations. It is useful when the task depends on files, packages, generated artifacts, or persistent execution state rather than reasoning over prompt context alone. It is not automatically required for every model interaction.
It helps to separate the control plane from compute. The harness can coordinate model calls, tools, approvals, tracing, recovery, and run state. A sandbox can execute model-directed work against a filesystem and command environment. Keeping those responsibilities apart can let trusted infrastructure retain authentication, billing, auditing, review, and recovery responsibilities while execution happens in an isolated environment.
A sandbox is not a complete safety policy. The harness still needs to define which actions are allowed, which require approval, and which are prohibited. Security depends on the boundary between components, including what credentials each component can access and what the execution environment is permitted to do. A sandbox without appropriately scoped permissions does not, on its own, establish that an agent is safe.
How is the model different from the harness?
| Part | Primary role | Example in a coding task |
|---|---|---|
| Model | Interprets the provided context and returns a response or action request. | Requests a command after deciding that it needs more information about the repository. |
| Harness | Supplies context and tools, routes requests, applies workflow rules, returns results, and tracks the run. | Checks the request against permissions, runs the command through the configured tool, and passes its output back to the model. |
| Execution environment | Runs actions against available files, commands, packages, or services. | Provides the workspace in which the requested command runs and may change files. |
The separation matters when diagnosing a failure. The model may request an unsuitable action; the harness may provide incomplete context or route a request incorrectly; or the execution environment may lack a needed file or package. These are different failure points, even though the user experiences them as one agent run.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Which runtime approach should an application use?
OpenAI’s documentation describes three approaches that place orchestration and state responsibilities differently. They are examples of runtime choices, not a universal ranking. The right balance depends on how much control an application needs and whether its tasks require persistent, isolated compute.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
| Approach | Who manages orchestration? | State and execution | Useful when |
|---|---|---|---|
| Agents API | OpenAI provides a managed Codex harness. | OpenAI manages state and infrastructure for longer-running work. | An application wants a managed runtime for longer-running coding tasks. |
| Agents SDK | The application controls deployment, storage, approvals, and runtime integration; the runner handles the loop and handoffs. | State and execution are integrated with the application’s runtime choices. | The application needs more control over deployment and integration while using a runner for the loop. |
| Responses API used directly | The application builds more of the integration itself around direct model calls. | The application’s implementation takes on more responsibility for history, chaining, and the surrounding workflow. | The application wants to construct more of its own agent integration. |
Before choosing, decide where orchestration, session state, tool execution, compute, approvals, credentials, and audit records should live. Then assess whether the task needs files, shell commands, packages, persistent artifacts, or resumable work. A runtime that makes orchestration convenient may leave less of the workflow under application control; a more direct integration can provide control while requiring the application to implement more of the surrounding system.
Best Value
What makes a coding-agent workflow easier to trust?
The harness is where an application’s workflow policies meet the model’s action requests. Useful engineering choices include making relevant repository context accessible, exposing a clear and appropriately scoped set of actions, preserving the state needed for the task, and making risky operations subject to permissions or review. The result should also be possible to inspect and check.
OpenAI’s account of its own 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 describes enforcing architectural invariants while leaving implementation choices open. These are practices from that workflow, not guarantees that every team should adopt an identical process.
For a coding task, the practical question is not simply whether the model can generate code. It is whether the complete system gives it enough relevant context, exposes suitable actions, applies appropriate boundaries, and provides a way to inspect the resulting changes.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →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.
