An agentic framework provides building blocks for developing agents. An agent harness provides the operating environment that runs those agents with tools, boundaries, state, and observability.
Four layers, four jobs.
The vocabulary overlaps because products often span more than one layer. The clearest comparison is to ask what the thing primarily helps you do: reason, implement, configure, or operate.
AI model
Generates responses, interprets context, and chooses a next action. It is the reasoning engine, not the complete agent system.
Agentic framework
Provides code-level primitives for tools, workflows, memory, state, events, and coordination when developers build agents.
Agent builder
Defines instructions, roles, tools, models, knowledge, and workflows through code, configuration, or a visual interface.
Agent harness
Runs the loop, assembles context, executes tools, applies permissions, manages workspaces, and records what happened.
Code harness
Adapts the harness to repository work: files, commands, tests, Git changes, branches, worktrees, and code review.
Agentic team
Connects a Director and specialist workers through visible delegation, shared outcomes, and controlled handoffs.
Detailed comparison.
A product may combine these responsibilities, but each row helps identify the capability you actually need.
| Capability | AI model | Agentic framework | Agent builder | Agent harness |
|---|---|---|---|---|
| Primary user | Agent runtime or application | Developer | Developer or operator | Operator and agent team |
| Core artifact | Generated response | Agent code and workflow | Configured agent or team | Running session and result |
| Tool definitions | May request tool calls | Defines interfaces and handlers | Assigns tools to agents | Validates and executes calls |
| Permissions | No operating boundary by itself | Can implement policy hooks | Can express intended access | Enforces access at runtime |
| Workspace | Receives supplied context | Can integrate with an environment | Configures resources | Maintains files, state, execution, and history |
| Observability | Model input and output | Tracing integrations | Configuration and status | Workers, actions, changes, tokens, cost, reports |
Which layer do you need?
Start from the gap in your current system. If you already have agent logic but cannot control or inspect its execution, another development abstraction is unlikely to solve the operating problem.
When you are building logic
You need code-level control over workflows, state transitions, tool adapters, memory, or application integration.
When you are shaping roles
You need a faster way to define agents, teams, instructions, knowledge, models, and the tools assigned to each role.
When you are operating work
You need agents to execute with controlled access, durable context, visible delegation, recoverable state, and an audit trail.
Where Baux sits.
Baux combines the builder and harness layers for local-first agent teams. You shape a Director and workers, give every role exact tools and a model, then operate the team in a workspace that keeps delegation, files, Git changes, usage, and results visible.
- Ready-made and custom agent team patterns
- Per-agent models, personas, and tool access
- One-shot routing to selected teammates or groups
- Serial or parallel delegation scheduling
- Project-contained worker directories and Git-aware execution
- Visible worker sessions and tool calls
- Tokens, cost, changed files, and final reports
Agentic framework FAQ.
Common questions about how the development and operating layers relate.
What is an agentic framework?
An agentic framework is a development toolkit that provides primitives for model calls, tools, state, workflows, memory, and multi-agent coordination.
What is the difference between an agentic framework and an agent harness?
A framework helps developers implement agent logic. A harness is the runtime and workspace that operates agents with context, tools, permissions, execution, and observability.
Do I need both a framework and a harness?
Sometimes. Teams building custom agent logic may use a framework underneath and a harness to operate it. Teams using ready-made agents may only interact with the harness and builder layers.