An agent system is software in which a language model selects the next action at runtime within an engineered harness. It may choose a tool, a response, or another reasoning step. A fixed pipeline that calls a model is not necessarily an agent: the defining feature is model-directed control flow. section 1.1
The model decides; the harness supplies context, memory, permissions, budgets, and stopping conditions. A feedback loop turns each result into a new observation until the task completes or a stopping rule fires. An agent system can contain one agent or several collaborating agents. section 1.1
Reliability depends on designing for uncertainty rather than assuming a fixed script:
- Output varies: the same prompt can produce different responses. The model can hallucinate or misinterpret instructions. section 1.5 section 1.3.1
- Behavior varies: the model selects its strategy at runtime and may loop or drift. Bound its decisions and define when to stop. section 1.5 section 1.3.1
- The environment changes between observations. Context, task records, retrieved evidence, and tool outputs may disagree, so state needs persistence, provenance, and attention policies. section 1.5 section 1.3.1
Architecture also allocates resources: the model spends, and the harness budgets. Evaluate designs for reliability and latency alongside token leverage—the ratio of output quality to token cost. Unbounded attention can bloat context, while unconstrained tool use increases the blast radius of mistakes. section 1.5 section 1.3.1
A coding agent misses project knowledge
A coding agent asked to add a controlled rollout created a second rollout utility, even though the repository already had one. The result looked reasonable in isolation but introduced a mechanism that could drift from the authoritative utility. Preventing that requires more than a prompt: the architecture must decide what knowledge the agent retrieves, what it may change, how its work is reviewed, and what evidence counts as complete. section 1.0
Agent design can be organized by function—perception, memory, reasoning, action, reflection, collaboration, and governance—and by execution topology, such as a chain, route, parallel fan-out, loop, hierarchy, or orchestration. These dimensions distinguish what the agent does from how it runs. section 1.1
