agent washing: the one question that exposes fake ai agents - By Sourav Mishra (@souravvmishra)
does it act on its own or only when you push a button? how to spot real autonomous agents.
"we use ai agents" is the new "we use ai." half the time, it's just a simple workflow doing one llm step. i call that agent washing.
in this post, i, sourav mishra, will give you the one question to expose fake agents. we'll see how to spot real ones and why this matters for your design and security. plus, i'll share a real pattern for building one.
the test: initiative vs trigger
the test: does it act on its own, or only when you hit run?
real agents run in loops: observe, decide, act, reflect. they pick tools on the fly based on what just happened.
if you can draw the whole flowchart before it even runs, you just have a workflow. the llm might fill in some blanks (like "approve or reject"), but it doesn't control the steps.
so when someone says they have an agent, ask: can it change its strategy based on the last tool's output?
if the answer is no, don't pay extra for it. if you're building one, use a real loop with tool-calling and a stop condition. here is a concrete pattern for that.
this difference isn't just wordplay. workflows are easy to secure and budget for. agents are super flexible but can be dangerous. messing this up ruins your design and security. i explain this more in agents vs workflows.
what agent washing looks like
agent washing is basically marketing something as an "ai agent" when it acts like a fixed script. no real autonomy. no dynamic tool choice.
for example, a "support agent" that is just a form sending data to an llm for a template response. or a "research agent" that always searches, summarizes, and returns.
the llm is working, but the path is set in stone. you draw it once, and it never changes. that's a workflow with an llm, not an agent.
real agents can loop back, try new tools, or change plans based on what they see. so again, just ask: can it change strategy? if yes, it's an agent. if no, it's a workflow. that's totally fine, just call it what it is.
why this matters
security and audits. workflows have fixed paths. you can lock down every single step easily.
agents, though, can use tools in unpredictable ways. they are way harder to audit and easy to abuse (like runaway api costs). if you buy a workflow thinking it's an agent, your security model is off.
costs and expectations. workflows have limits. agents can run loops until they hit a stop condition.
this means costs and waiting times are less predictable with agents. if you confuse the two, your budget planning goes out the window.
building the right thing. if you need open-ended research or support, a workflow is too limited.
but if you just need strict approvals or forms, an agent is way too risky. know the difference so you build (or buy) the right tool for the job.
how to build a real agent
if you want to build a real agent, you need: a single loop, tools, an llm, and a stop condition.
the agent observes the state, decides what to do, acts on it, updates its memory, and repeats. it stops when it reaches a limit.
i use the vercel ai sdk and add rules like stopWhen: stepCountIs(N). this stops it from running forever. you can see the full setup here: building an agentic chatbot.
for production, give tools the minimum access needed. add a human-in-the-loop for big actions. see production-ready agents for more tips.
can langgraph be used for real agents?
yep! it's agentic if you use loops and let it pick tools dynamically. otherwise, it's just a workflow. same rule applies: can it change strategy based on the last output?
key takeaways
- agent washing: calling a fixed workflow an agent. just ask: can it change strategy? if no, it's a workflow.
- real agents: observe, decide, act, reflect. they pick paths at runtime. workflows: fixed steps, llm just fills in blanks.
- why it matters: workflows are safer and easier to budget. agents are flexible but risky.
- building an agent: use a loop, tools, llm, and a stop condition. check out my agentic chatbot guide.
written by sourav mishra. full stack engineer, next.js and ai.
frequently asked questions
q: what is agent washing? it's marketing a fixed script as an "ai agent." if it can't change its plan based on tool outputs, it's just a workflow.
q: how do i spot a real agent? see if it chooses different paths based on results. if it just runs a fixed sequence when triggered, it's a workflow.
q: can langgraph build real agents? yes, if you use loops and dynamic tool choices. otherwise, it's a workflow. see agents vs workflows.
q: where is a good agent example? i wrote a guide on building an agentic chatbot with vercel ai sdk. it covers loops, tools, and safety limits.