The Agent Loop: Observe → Plan → Act → Verify

{
  "byte_id": "agentic_dev_001",
  "title": "The Agent Loop: Observe → Plan → Act → Verify",
  "level": "foundation",
  "domain": [
    "agentic-development",
    "llm-agents"
  ],
  "intent": "Understand the minimal mental model of an AI agent so you can explain it clearly and design it reliably.",
  "core_idea": {
    "one_liner": "An agent is not 'a smart chat' — it is a loop that repeatedly reads the world, decides what to do next, executes actions via tools, and checks results.",
    "why_it_matters": [
      "Prevents magical thinking: the agent needs data, actions, and verification.",
      "Makes failures debuggable: you can see which step broke (observe/plan/act/verify).",
      "Turns vague tasks into predictable workflows."
    ]
  },
  "loop_steps": [
    {
      "step": "Observe",
      "goal": "Collect the minimum facts needed to make a decision.",
      "typical_inputs": [
        "user request",
        "system state",
        "retrieved knowledge (RAG)",
        "tool outputs",
        "constraints/policies"
      ],
      "rules_of_thumb": [
        "If the agent is unsure, it must fetch/ask rather than guess.",
        "Prefer structured observations (tables, JSON, logs) over prose."
      ],
      "failure_modes": [
        "Hallucination due to missing facts",
        "Using stale knowledge",
        "Ignoring constraints (security/policy/business rules)"
      ]
    },
    {
      "step": "Plan",
      "goal": "Choose the next best action sequence with clear success criteria.",
      "typical_outputs": [
        "short plan",
        "decision points",
        "required tools",
        "expected outputs"
      ],
      "rules_of_thumb": [
        "Plan should be short and actionable (3–7 steps).",
        "Define a stop condition: what does 'done' look like?"
      ],
      "failure_modes": [
        "Overplanning (too long, never executes)",
        "Underplanning (jumps into actions blindly)",
        "No success criteria (can't verify)"
      ]
    },
    {
      "step": "Act",
      "goal": "Execute the plan using tools (APIs, DB queries, file edits, webhooks) safely and idempotently.",
      "typical_actions": [
        "call tool",
        "write draft",
        "generate JSON",
        "run query",
        "create ticket"
      ],
      "rules_of_thumb": [
        "Prefer deterministic actions (tool calls, structured outputs).",
        "Make actions idempotent where possible (safe retries)."
      ],
      "failure_modes": [
        "Tool errors/timeouts",
        "Repeating actions (duplicate side effects)",
        "Wrong tool selection or wrong parameters"
      ]
    },
    {
      "step": "Verify",
      "goal": "Check that the result matches the success criteria; if not, loop back.",
      "verification_methods": [
        "compare with expected output",
        "sanity checks",
        "cross-check sources",
        "unit tests / evals"
      ],
      "rules_of_thumb": [
        "Verification must be explicit (not 'seems ok').",
        "If verification fails, the agent should produce a diagnosis and next step."
      ],
      "failure_modes": [
        "No verification (silent wrong answers)",
        "Fake verification (claims it checked but didn’t)",
        "No fallback strategy"
      ]
    }
  ],
  "when_to_use": [
    "Any multi-step task: research, troubleshooting, content generation, data cleansing, support workflows.",
    "When correctness matters and you need traceability."
  ],
  "when_not_to_use": [
    "Pure creativity with no correctness constraints (poems, brainstorming).",
    "Ultra-simple one-shot answers where tools/verification add unnecessary overhead."
  ],
  "micro_example": {
    "scenario": "SAP MDG BP replication is slow. What should we do?",
    "observe": [
      "Get queue/monitor data (DRF, web service logs, retries, backlog size).",
      "Check if delay is consistent or spikes.",
      "Confirm what 'slow' means (minutes vs hours)."
    ],
    "plan": [
      "Identify bottleneck category (queue backlog vs technical failures vs downstream).",
      "Run checks in order (quick wins first).",
      "Propose mitigation + long-term fix."
    ],
    "act": [
      "Pull queue metrics and error samples.",
      "Generate a short RCA candidate list + confirmation steps."
    ],
    "verify": [
      "Do metrics improve after mitigation?",
      "Did we confirm root cause with logs?",
      "Are we sure it's not downstream capacity?"
    ]
  },
  "practical_checklist": [
    "Do we have enough observations to act, or are we guessing?",
    "Is there a short plan with success criteria?",
    "Are actions safe to retry (idempotent)?",
    "Do we verify with something real (logs/tests/second source)?"
  ],
  "pitfalls": [
    "Skipping Observe → hallucination",
    "Skipping Verify → confident wrong output",
    "Using RAG as 'truth' without checking freshness/version"
  ],
  "teach_it_in_english": {
    "simple_explanation": "Think of an agent like a junior engineer with a checklist: it looks at the situation, decides the next steps, uses tools to do work, and then checks if the result is correct. If not, it loops and improves.",
    "one_sentence_definition": "An AI agent is a tool-using decision loop with explicit verification."
  },
  "tags": [
    "agent-loop",
    "react",
    "plan-execute",
    "verification",
    "reliability"
  ],
  "meta": {
    "schema": "dkharlanau.dataset.byte",
    "schema_version": "1.1",
    "dataset": "agentic-bytes",
    "source_project": "cv-ai",
    "source_path": "agentic-bytes/agentic_dev_001.json",
    "generated_at_utc": "2026-02-03T14:33:32+00:00",
    "creator": {
      "name": "Dzmitryi Kharlanau",
      "role": "SAP Lead",
      "website": "https://dkharlanau.github.io",
      "linkedin": "https://www.linkedin.com/in/dkharlanau"
    },
    "attribution": {
      "attribution_required": true,
      "preferred_citation": "Dzmitryi Kharlanau (SAP Lead). Dataset bytes: https://dkharlanau.github.io"
    },
    "license": {
      "name": "",
      "spdx": "",
      "url": ""
    },
    "links": {
      "website": "https://dkharlanau.github.io",
      "linkedin": "https://www.linkedin.com/in/dkharlanau"
    },
    "contact": {
      "preferred": "linkedin",
      "linkedin": "https://www.linkedin.com/in/dkharlanau"
    },
    "canonical_url": "https://dkharlanau.github.io/datasets/agentic-bytes/agentic_dev_001.json",
    "created_at_utc": "2026-02-03T14:33:32+00:00",
    "updated_at_utc": "2026-02-03T15:29:02+00:00",
    "provenance": {
      "source_type": "chat_export_extraction",
      "note": "Extracted and curated by Dzmitryi Kharlanau; enriched for attribution and crawler indexing."
    },
    "entity_type": "agentic_byte",
    "entity_subtype": "level:foundation",
    "summary": "Understand the minimal mental model of an AI agent so you can explain it clearly and design it reliably."
  }
}