Calling software an 'agent' sounds impressive, but does that label actually tell you what it can do?
Does calling a system an agent tell us anything concrete about its actual capabilities?
This explores whether the word 'agent' points to a specific set of abilities, or whether it's a loose label that hides big differences between systems.
This explores whether the word 'agent' tells you what a system can actually do, or whether it hides large differences between systems that all go by the same name. The collection's short answer: not much on its own. Several frameworks drop the yes-or-no label and use a spectrum instead. OpenAI's 2023 framework rates 'agenticness' on four separate dials: how complex the goals are, how messy the environment is, how well the system adapts, and how long it runs without help. A system can score high on one dial and low on another Should we define AI agents as a spectrum rather than a category?. Nielsen Norman Group goes further. In their glossary, autonomy isn't a property of the AI at all. It's set by deployment choices: what the system is allowed to do, when it has to stop, and when a human reviews its work. Even 'human-in-the-loop' only counts as oversight if the reviewer has the information, time, expertise and authority to step in How much autonomy does an agentic AI system actually have?.
A second idea follows from this: much of what we call an agent's capability isn't in the model. Research on agent 'harnesses' (the software around the model that handles memory, reusable skills and structured interaction) argues that reliability comes from this scaffolding rather than from a bigger model Where does agent reliability actually come from? Where does agent reliability actually come from?. Dubberly's layered model shows the same thing from a design angle. A chatbot, a chatbot with a system prompt, a chatbot with document retrieval (RAG), and a system of supervisor agents could all be called 'agents', but each layer adds inputs and oversight that change what the system can do How should agent capabilities be organized into functional layers?. Even the medium matters. Agents that reason in code, which can be run, inspected and can hold state, behave differently from agents that reason in plain prose Can code serve as the operational substrate for agent reasoning?.
This is why the usual way of measuring agents tells you so little. Two agents with the same task-success rate can differ hugely in efficiency, reliability and readiness for real use How should we measure agent system performance beyond task success?. Memory systems need to be tested stage by stage (storing, extracting, retrieving, maintaining) to find which part actually failed How should we actually evaluate agent memory systems?. One analysis of 960 real job workflows found that agents win benchmark contests but fail long, multi-step professional work. Its authors call this a gap in how we measure, not in what the models can do: the field has been measuring contests rather than work Why do agent benchmarks not predict real economic value?. Adding more agents doesn't settle it either. Across 180 setups, what predicted success was how well the setup's structure fit the task, not how many agents there were. Past about 45% single-agent accuracy, coordination stopped helping, and the way agents were connected changed how much errors multiplied by 4 to 17 times When does adding more agents actually help systems?.
The unexpected part is what 'agent' does tell you: it's a warning about risk more than a promise of ability. Once a system can touch its environment through memory, retrieved documents and tool calls, a filter that checks one output at a time can't contain it. Safety has to come from limiting what it can reach Can a model-level filter truly contain an agent with environment access?. And once agents start buying, deploying and transacting, the main limit is no longer intelligence. It's infrastructure: identity, delegation, audit trails and accountability Does agent capability matter more than coordination infrastructure?. So when something is called an agent, the useful questions are what it's allowed to touch, what scaffolding surrounds it, and how it was tested, not how smart it is.
Sources 12 notes
OpenAI's 2023 framework measures AI agenticness across goal complexity, environmental complexity, adaptability, and independent execution as separate dimensions. This approach avoids false binaries while capturing systems that score high on some axes but low on others.
Nielsen Norman Group's glossary reframes 'agentic' from a binary label to a spectrum of autonomy determined by permissions, stopping rules, and human-review conditions rather than the agent mechanism itself. Human-in-the-loop counts as oversight only when reviewers have information, time, expertise, and authority to actually intervene.
Applied AI research shows capability shifts from model weights to external structures like memory and skills. However, reusable skills bundle executable code and system reach, creating security costs that traditional lifecycle inspection cannot catch when attacks compose across multiple skills.
Research shows reliable LLM agents externalize three cognitive burdens—memory (state persistence), skills (procedural components), and protocols (structured interaction)—into a harness layer rather than relying on model scale alone. The harness unifies these externalities and eliminates the need for the model to solve the same problems repeatedly.
Dubberly proposes a two-axis stack model: a capability axis (chatbot → system prompt → RAG) adding inputs that extend what LLMs can act on, and a scale axis (monitoring agents → supervisors → orchestrator) aggregating oversight. He cautions that ubiquitous agent hierarchies raise design risks alongside capability gains.
Show all 12 sources
Research shows code uniquely enables agent reasoning, action, and verification by being simultaneously executable, inspectable, and stateful. This unified code-centered loop improves reasoning and verification together compared to natural-language or prose-based approaches.
Single task-success metrics obscure how agents achieve results across memory, context, and verification layers. Research shows identical success rates can mask enormous differences in efficiency, reliability, and deployment readiness—requiring harness-level benchmarks that measure trajectory, memory hygiene, and verification costs.
Decomposing memory into storage, extraction, retrieval, and maintenance stages exposes design trade-offs and failure modes that task-success metrics completely hide. Module-by-module evaluation across 12 systems shows which component actually failed rather than just whether the task succeeded.
ALE's analysis of 960 real occupational workflows shows agents excel at abstract contests but fail long-horizon professional tasks. The gap is not model capability but benchmark design—the field optimizes what it measures, and it has measured contests rather than work.
Across 180 configurations, three dominant effects predict multi-agent success: tool-coordination trade-offs harm complex tasks, coordination stops helping above 45% accuracy, and topology choice controls error amplification by 4–17×. Architecture-task alignment, not agent count, determines outcomes.
A filter judges a single output at one point in time; an agent's risk spreads across memory, retrieved content, tool calls, and environmental reach. Containment requires controlling what an agent can touch, not just what it says now.
Once agents move beyond simple API calls to purchasing, deploying, and transacting with real consequences, the bottleneck shifts from model capability to whether they can coordinate reliably, maintain accountability, and produce auditable evidence. Infrastructure—identity, delegation, attestation, and audit trails—matters more than marginal improvements to reasoning.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Holistic Agent Leaderboard: The Missing Infrastructure for AI Agent Evaluation
- Towards a Science of AI Agent Reliability
- Towards a Science of Scaling Agent Systems
- From Model Scaling to System Scaling: Scaling the Harness in Agentic AI
- AI Agents That Matter
- Agentifying Agentic AI
- Survey on Evaluation of LLM-based Agents
- Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering