Coding-agent workspaces, explained

Agentic

Understand the pieces behind a coding agent: what runs it, what it can do, and what guides its work. Follow the connections, then explore each building block.

Anatomy of a coding-agent workspace

Ten building blocks. Follow how context, model responses, and tool results drive the next step.

ProjectOwned by this project
Execution / feedbackCapability connection / exposureGuidance / contextCompatible component deliverySelect a block or colored dot to explore its contents.

Harness

The harness is the surrounding software. An agent pursues a goal inside it; the model supplies the reasoning.

  • Harness sends context & tool results to Model
  • Model returns text or tool requests to Harness
  • Plugins supply components to Harness
  • Harness connects to MCP servers
  • Harness triggers Hooks
  • Harness runs Agent
  • Harness executes authorized calls to Tools

Harness / README.md

Example template

How Harness fits in

The harness is the surrounding software. An agent pursues a goal inside it; the model supplies the reasoning.

How an agent harness works

Project workspace harness

Runs a coding agent in a local project workspace.

Responsibilities

  • Load the task, project instructions, and relevant context.
  • Send the conversation and available tools to the model.
  • Execute permitted tool calls and return their results to the model.
  • Feed tool results and errors into the next model call so the agent can assess progress.
  • Continue until the task is complete, needs input, or reaches a configured limit.

Runtime choices

SettingExample
WorkspaceCurrent project directory
ToolsRead files, edit files, run project checks
PermissionsWorkspace edits allowed; publishing requires approval
LimitsSet a time or cost budget for each run
Session stateSave progress and tool results for resuming work

Completion

Report the changed files, checks run, and any remaining work. Keep credentials outside prompts and logs.

A starting point for your project. Adapt the names, commands, and configuration to your agent app.