## Definition **Agentic feedback taxonomy** is the classification of the feedback signals injected into a planner prompt — or into subsequent LLM calls — to correct and improve an agent's plan. Lanham identifies three types in *[[AI Agents in Action - Micheal Lanham]]*: **corrective** (what went wrong), **suggestive** (what to try instead), and **epistemic** (what the agent was missing). Together, they replace post-hoc evaluation with structured, reusable improvement signals that can be carried across planning iterations. ## The Three Types ### Corrective Feedback Points to a specific error in the last plan or execution step. It does not propose an alternative — it only names the failure: > "Step 3 assumed `get_wikipedia_page` returns plain text, but it returns HTML." Injected as a constraint in the next planner call, corrective feedback narrows the space of plans the planner should generate. ### Suggestive Feedback Proposes a remedy without mandating it. The planner retains latitude to incorporate the suggestion or route around it: > "Consider adding a parsing step after `get_wikipedia_page` to strip HTML tags before passing content downstream." Suggestive feedback is useful when the evaluator knows *a direction* but not the canonical solution — it guides without over-specifying. ### Epistemic Feedback Addresses a gap in the agent's knowledge rather than a flaw in its reasoning: > "The agent did not know that `save_file` requires an absolute path; relative paths silently fail." Epistemic feedback is the most durable of the three: it can be promoted into the agent's standing system prompt or AGENTS.md file so that future plans never repeat the same ignorance. ## How Feedback Is Injected In Lanham's Nexus planner, all three types are concatenated as a `[FEEDBACK]` block in the planning prompt template alongside `[GOAL]` and `[AVAILABLE FUNCTIONS]`. The planner is instructed to incorporate the feedback when constructing the next plan. This keeps feedback in-prompt and stateless — no external memory required for single-session use. For multi-session persistence, epistemic feedback in particular should migrate to [[Architecture Decision Record]] or [[AGENTS.md Convention File]] so it survives context resets. ## Relation to Other Agent Patterns - **[[Reflection]]** is self-generated corrective feedback — the agent critiques its own output. Agentic feedback taxonomy governs *externally injected* feedback, which may or may not be self-generated. - **[[Builder-Critic Pattern]]** produces corrective and suggestive feedback as outputs of the critic role, which are then fed back to the builder — a structured instance of this taxonomy. - **[[Agent Planning]]** is the consumer of the feedback: the planner prompt is the point of injection. ## Why Typed Feedback Matters Mixing all feedback into a single untyped block tends to produce plans that address one signal while ignoring others. Typing feedback separates "here is what broke" from "here is what to try" from "here is what you didn't know" — each answered with a structurally different planning response. It also makes feedback auditable: you can inspect which type was dominant and decide whether to promote it to persistent memory. ## Related - [[Reflection]] - [[Builder-Critic Pattern]] - [[Agent Planning]] - [[Prompt Chaining]] - [[Architecture Decision Record]] - [[AGENTS.md Convention File]] ## Sources - [[AI Agents in Action - Micheal Lanham]]