An agent is a software system in which a language model chooses the next action at runtime within an engineered harness. The model interprets observations and selects a tool, a response, or another reasoning step. A fixed pipeline that calls a language model is not necessarily an agent: the defining feature is model-directed control flow. section 1.1
The model supplies flexibility; the harness supplies context, memory, permissions, budgets, and stopping conditions. A feedback loop turns each result into the next observation until the task completes or a stopping rule fires. A system can contain one agent or several collaborating agents. section 1.1
Agent architecture can be organized around seven functions: perception, memory, reasoning, action, reflection, collaboration, and governance. Patterns are reusable design choices within those functions; execution topologies describe how choices run, such as a chain, route, parallel fan-out, loop, hierarchy, or orchestration. section 1.1
A general-purpose agent can discover Skills, use a command line or browser, inspect files, query databases, and work through errors. Stronger models and platform primitives may eliminate custom handoffs, memory layers, or reasoning scaffolds. If a simpler baseline passes on rerun, it can replace a more elaborate composition. section 10.5
Capability does not remove the need for executable boundaries. The model may propose a command, but the platform executes it and returns bounded results. More context alone does not authorize external actions, commit shared state, make retries idempotent, or prove that a business result exists. Permissions, durable state, recovery, and external acceptance become more explicit as consequences increase. section 10.5
From code review to customer escalation
A coding agent and a customer-escalation agent face similar design decisions: which evidence to load, which actions to permit, and when a human must approve the result. The task changes, but the harness must still bound the model's choices. section 1.1
Table 10.5 Interface explicitness follows consequence and distance
|
Workload
|
Default
composition stance
|
|
Short,
single-user, read-only or reversible exploration
|
Use a
general agent with standard tools; make dangerous actions and the final
artifact explicit
|
|
Multi-step
work across agents, sessions, or owners
|
Add
lightweight contracts for run identity, artifacts, state, and handoff
ownership
|
|
Multi-tenant
production work with money, identity, regulated data, or irreversible effects
|
Use
contract-first boundaries for authority, provenance, durable state,
idempotency, recovery, and external acceptance
|
Compare interface boundaries for reversible exploration, work spanning agents or owners, and consequential multi-tenant production work. section 10.5
