Tool calling
GLM-5 models on OpenRouter support tools and tool_choice, the building blocks of an agent loop.
use-cases/agents
Use case · agentic engineering
Building agents with GLM-5
Long-horizon agent work is the main design goal of the GLM-5 family. GLM-5.1 was built to keep improving a solution over hundreds of rounds and thousands of tool calls, and GLM-5.3 extends that with a million-token context. This guide covers model choice, design and how to use our chat to plan an agent.
Which modelGLM-5.3For production agents, use GLM-5.3 (or GLM-5.3 Prime for speed) through the API with tool calling. GLM-5.3 Flash is a strong, much cheaper option for high-volume or vision-based agents. In our chat, use GLM-5.3 to design and debug agent prompts.
§ 01What it does well
Why GLM-5 fits agent work
Long horizons
Z.ai designed the family for tasks that run for many steps: plan, act, observe, revise.
Huge context
About one million tokens lets an agent keep a whole project and its history in view.
Cheap input
Agents resend context on every step; GLM's low input prices keep that affordable.
§ 02Prompts
Copy-ready prompts
Open any prompt in the chat, then replace the bracketed parts with your own material.
Design an agent
Help me design an AI agent that [goal]. List the tools it needs (name, inputs, outputs), the main loop, how it decides it is done, and the three most likely failure modes with mitigations.
Write a system prompt
Write a system prompt for an agent that [task]. Include its role, the tools it may call, rules for when to ask the user, and the exact output format for its final answer.
Debug an agent run
Here is a log of my agent's run where it went wrong. Find the step where it first made a bad decision, explain why, and suggest a change to the prompt or tools. [paste log]
§ 03Tips
Do this, avoid that
Do
- Keep tools small with clear input schemas.
- Set a maximum number of steps.
- Log every tool call and result.
- Use a low reasoning effort for routine steps.
Avoid
- Do not give agents unrestricted write access.
- Do not skip human approval for risky actions.
- Do not resend unchanged huge context without need.
- Do not trust tool outputs blindly.
§ 04FAQ
Questions
Q01Which GLM model is best for agents?
Z.ai positions GLM-5.3 for long-horizon agent tasks. GLM-5.3 Prime is the faster variant, and GLM-5.3 Flash is the cheap, multimodal option.
Q02Can I run an agent in your chat?
No. The chat has no tools. Use it to design prompts and debug logs, then run the agent through the API.
Q03Do GLM-5 models support function calling?
Yes. On OpenRouter they list tools and tool_choice among their supported parameters.
§ 05More
Other use cases
Try it now
Ask GLM-5.3 something real.
No account, no card, no install. Type a question and watch GLM-5.3 Flash answer in seconds — switch to GLM-5.3 for the hardest problems.
Open the free chatFree daily limits apply so the chat stays open for everyone.