An agent harness is the software layer that bounds a language model’s decisions while it directs an agent’s next action. The model chooses a tool, response, or reasoning step; the harness supplies context, memory, permissions, budgets, and stopping conditions. A feedback loop turns results into observations until the task finishes or a stopping rule fires. section 1.1
Unlike a fixed pipeline, an agent loop leaves action selection to the model at runtime. The harness constrains that decision space: it determines what the agent can read or change, where durable state lives, and what evidence counts as completion. Loading a document or spawning a subagent spends context, compute, or action opportunities, so these choices need budgets as well as capabilities. section 1.2
A harness can support planning, delegation, and filesystem-backed context, but it does not choose the collaboration pattern. The application still decides how responsibilities are split and whether work runs as a chain, an independent batch, a review loop, or a hierarchy. A long-running system may also need a runtime with checkpoints and recovery. section 8.8
Delegation with Deep Agents
Deep Agents gives a lead agent a task tool and lets the application define subagents with distinct instructions, descriptions, and tools. The lead model decides when to delegate; each subagent works in a separate context and returns its result. The application remains responsible for the overall collaboration topology. section 8.8
Table 8.7 Implementation layers and where collaboration control lives
|
Implementation layer (examples)
|
Where control lives
|
What it supplies
|
Topologies
|
|
Orchestration runtime (LangGraph)
|
State graph or phase controller
|
Durable state, recovery, routing
|
All four
|
|
Agent framework (LangChain, OpenAI Agents SDK)
|
Current agent or agent loop
|
Model-tool loop, handoffs, routing
|
Depends on the application
|
|
Agent harness (Deep Agents)
|
Lead agent, middleware, and subagent tool
|
Planning, filesystem context, delegation
|
Hierarchy with embedded chain, parallel, or review
|
|
Team or process framework (CrewAI, MetaGPT)
|
Crew, manager, or process
|
Defined roles and process
|
Chain, parallel, hierarchy
|
|
Conversational framework (AutoGen/AG2, CAMEL)
|
Agents and message flow
|
Interactive exchange
|
Chain and review
|
|
Protocol layer (MCP, A2A)
|
Outside the task topology
|
Tool/data access or task/result exchange
|
None by itself
|
Compare where control lives in a harness, an orchestration runtime, an agent framework, and other collaboration layers. section 8.8



