Stop calling every LLM interaction an "agent." A new deep-dive from a Japanese founder who runs 47 concurrent Claude Code sessions clarifies the critical architectural distinction: chatbots answer, workflows follow scripts, and agents act. The piece, published on DEV.to on November 19, 2026, leans heavily on Anthropic’s engineering definitions to map out exactly where your autonomy ends and the model’s begins.
The Triad: Chatbot, Workflow, Agent
The core confusion stems from lumping three distinct systems under the "AI" umbrella. Anthropic’s December 2024 article "Building effective agents" provides the taxonomy: a chatbot replies with text while you do the work; a workflow orchestrates tools through predefined code paths; and an agent is a system where LLMs dynamically direct their own processes and tool usage. The author, who has run a company in Japan for 22 years, emphasizes that agents are "typically just LLMs using tools based on environmental feedback in a loop." This loop is what separates true agency from mere automation.
Inside the Loop: Gather, Act, Verify
Claude Code’s documentation explicitly names the agent’s internal cycle: gather context, take action, and verify results. Unlike a chatbot that stops at generation, an agent checks its own work against ground truth. For example, if tasked with fixing failing tests, the agent runs the suite, reads error output, edits files, and then re-runs the tests to confirm the fix. This verification step is the "ground truth" Anthropic describes, allowing the agent to self-correct without human intervention until it hits a stopping condition or a permission barrier.
The Five Patterns Before You Build an Agent
Before deploying complex agentic systems, Anthropic recommends evaluating five production-proven patterns: prompt chaining, routing, parallelization, orchestrator-workers, and evaluator-optimizer. The author notes that Anthropic’s own advice is to find the simplest solution, often meaning no agent at all. If the steps are fixed, a workflow offers better predictability. Agents are only recommended when flexibility and model-driven decision-making are needed at scale, as autonomous nature introduces higher costs and the potential for compounding errors.
Safety Rails: Permissions and Hooks
Control is maintained through permission modes and hooks. Claude Code offers four modes: Manual (asks before edits), Accept edits (allows common filesystem commands), Plan (proposes without editing), and Auto (uses a classifier to block risky actions). The author details using .claude/settings.json to define deny rules that always win, such as blocking git push while allowing git commit. For deterministic control, hooks run shell commands at fixed points, such as PreToolUse, which can block a tool call by exiting with code 2 before execution.
Key Takeaways
- Agents dynamically direct their own processes, unlike workflows which follow predefined code paths.
- The core agent loop consists of gathering context, taking action, and verifying results against the environment.
- Anthropic advises starting with simple prompts or workflows before introducing the complexity and cost of agents.
- Remote actions like database writes or emails cannot be checkpointed and require strict permission rules.
- Hooks provide deterministic control to block risky commands before they execute, independent of the LLM's decision.
The Bottom Line
An agent is not a smarter chatbot; it is a loop you trust with one job. If you cannot write down the definition of "done" and the verification criteria, you do not need an agent—you need a better prompt.
Sources
DEV.to: AI agents explained for beginners