AI Agents
AI agents are language-model systems that use tools and environmental feedback to pursue goals across multiple steps. The model participates in deciding what to do next rather than only producing a single response. OpenAI source record · Anthropic source record
Boundary and terminology
Anthropic uses agentic systems as an umbrella term but distinguishes fixed workflows from agents: workflows follow predefined code paths, while agents dynamically control their process and tool use. OpenAI’s definition is narrower: the model must manage workflow execution, choose tools, detect completion, and recover or return control to the user. Anthropic source record · OpenAI source record
These definitions are compatible but not identical. Analysis: When comparing systems, it is safer to describe who controls execution and how much of the path is predetermined than to rely on the word “agent” alone.
Core structure
Across the practical guides, a minimal agent includes:
- a model that interprets state and selects a next step;
- instructions or policies that constrain behavior;
- tools that retrieve information or take action;
- observations from the environment;
- a loop that continues until completion, failure, escalation, or another stopping condition.
OpenAI explicitly groups the foundations as model, tools, and instructions. Anthropic describes the implementation as a model using tools in a feedback loop and stresses ground truth from tool results or code execution. OpenAI source record · Anthropic source record
ReAct is a research pattern that makes the reasoning–action–observation cycle explicit by interleaving task reasoning with actions and environmental observations. ReAct source record
Harness and deployment
The model loop operates inside an agent harness that assembles context, dispatches tools, maintains state, enforces controls, evaluates progress, and exposes operational traces. The New SDLC whitepaper adds deployment and production infrastructure to the conceptual agent stack. New SDLC source record (pp. 10–30)
Analysis: This means an agent should be identified by more than its model and prompt. Tool versions, permissions, memory, orchestration, evaluators, and deployment configuration can materially change its behavior.
When agents fit
Agents are most plausible when the path cannot be specified reliably in advance: the task involves ambiguous judgment, changing context, unstructured information, or an unpredictable number of steps. Fixed automation or a single augmented model call remains preferable when it can meet the requirement more cheaply and predictably. OpenAI source record · Anthropic source record
Useful environments also provide measurable outcomes, meaningful tool feedback, and opportunities for human review. The sources highlight coding and customer support because tests, system records, or user-defined resolutions can ground progress. Anthropic source record
Costs and failure modes
Autonomy can improve flexibility while increasing latency, cost, and the chance that errors compound across steps. Reliability therefore depends on evaluations, clear interfaces, stopping conditions, and safety and control mechanisms. Anthropic source record · OpenAI source record
Evidence status
The architecture and deployment advice here is supported by vendor engineering guides and a Google course whitepaper that report practical experience but not systematic comparative studies. ReAct contributes benchmark evidence for one reasoning-and-acting format, but its results are specific to the paper’s tasks and baselines. Confidence is medium for broad design recommendations and should be raised only with application-specific evaluations or independent studies.