Agent Architecture and Orchestration
Agent Architecture & Orchestration
Scope: agent planning, coordination, execution, and the systems that make agent work reliable.
Capabilities: 19
Agent engineering entry points
| Design question | Start here | Continue with |
|---|---|---|
| What environment lets the agent execute and verify work? | Agent Harness Engineering | Context assembly, tool dispatch, work state, and evaluation |
| What steps and transitions may the system take? | Agent Graph Engineering | Node contracts, state updates, guarded paths, and recovery |
| What selects work, checks it, and decides what happens next? | Loop Engineering | Inner turns, outer work cycles, scheduling, and acceptance |
These are complementary views of an agent system, with overlapping components. A bounded model/tool loop may be one node in a graph, while an outer controller selects graph runs and checks their outcomes. The harness provides the environment for that work. Choose the structure required by the task rather than adopting three separate frameworks.
For a first design, read the harness component map, the graph's worked ticket workflow, and the loop's incident-triage example. Then follow their links to the specialized capabilities below. Graphs here organize execution; knowledge graphs organize facts and relationships for reasoning and retrieval.