# An Agent Is Defined by Working Without Mid-Task Steering ## Claim A tool only counts as an agent if, once given requirements, it completes the task and hands back finished output without needing turn-by-turn human input — anything that still requires ongoing steering is an assistant wearing agent marketing. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds for classifying any tool labeled "agentic" — chatbots, copilots, workflow assistants — regardless of vendor claims. - A supervised agent with deliberate approval gates still counts as an agent; the test is whether steering is *needed* to get a correct result, not whether a human *chooses* to check in. - Breaks down for genuinely mixed-mode tools that alternate between agentic bursts and assistant-mode clarification — the label should probably track the dominant mode, not force a binary. ## Evidence Christopher Penn draws the line directly: "An AI agent or any agent is a piece of software that does something without you and returns its results to you," and the operational test follows immediately — "if you are still providing input and feedback and steering throughout the process, it's not an agent." He applies this to reject Microsoft's "Copilot agents" as SOPs, not agents, on exactly this basis. ## Links - [[3. Resources/Permanent Notes/An Agent's Default Persona Should Be Collaborator Not Assistant]] — contrasts: that note is about the disposition an agent should have once autonomous; this note is about the prior question of whether a tool qualifies as autonomous at all. - [[3. Resources/Permanent Notes/Route Judgment to the Expensive Model and Bulk Execution to the Cheap One]] — applies-in: cost-tiered planning only makes sense once a tool has cleared this autonomy bar; a steered assistant doesn't have a "planning" phase to route separately. - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — this definitional test is the entry criterion for any claim that a workflow has been "agentified." - [[3. Resources/Permanent Notes/Judgment at the End of an Agent Run Turns Every Late Failure Into a Restart]] — contrasts: the handover needs a point to interject before the run ends (Kemple). ## Tests / falsifiers - Take a tool marketed as an agent and remove the human from the loop entirely after the initial task is specified; if output quality or task completion collapses, the tool was an assistant with an agent label. - Track how often a "level 3+" agentic tool (per Penn's 5-level maturity ladder) still requires mid-task correction in practice — a high correction rate would undercut the clean binary this note assumes. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Readwise/Articles/Almost Timely News How to Connect an AI Agent to a Data Source]] — Christopher S. Penn, 2026-07-19; the agent/assistant litmus test, applied to reject "Copilot agents" as non-agentic.