How do you design a function-calling agent without writing glue code first?

Key facts

Tool type	Free visual builder for function-calling agents
Agent features	Agents, function calling, memory and orchestration are live
Tool definitions	JSON Schema for parameters, with descriptions the model reads
Loop controls	Max steps, timeouts and stop conditions
Models	30+ models testable from the live catalogue
Trace	Step-by-step view of model output, tool calls and results
Free plan	2 free AI models (plugsky-micro, plugsky-lite), no card required
Product status	Live (agents, function calling, memory)

TL;DR

• Define tools with JSON Schema and write descriptions for the model, not for humans.
• Set loop limits before you test, so a confused agent cannot run forever.
• Read the trace: most agent failures are schema or description problems, not model problems.
• Test the same agent on a small free model and a frontier model to see where quality breaks.
• Export the configuration and keep the loop logic in your own service for observability.

How it works, step by step

1. Open the AI agent builder and write the agent goal as a short, testable system prompt.
2. Add each tool: name, description and parameter schema, keeping parameters flat where possible.
3. Set loop guardrails — maximum steps, per-step timeout and a clear stop condition.
4. Choose a model and run the agent on a representative task.
5. Inspect the trace: which tool was called, with what arguments, and whether the result was used correctly.
6. Tighten descriptions and schemas where the model hesitated or chose the wrong tool.
7. Re-run until the trace is clean, then export the configuration and port the loop into your service.

Try it yourself

Open the AI agent builder →

Designing tools the model can actually use

Tool quality decides agent quality. Give every tool a precise name, a description that states when to use it and when not to, and a JSON Schema with sensible types, enums and required fields. Avoid overlapping tools that do almost the same thing; the model will oscillate between them and waste steps. Keep arguments flat: nested objects increase malformed-call rates. Return compact, structured tool results — a wall of text forces the model to re-parse what you already parsed in code. If a tool can fail, define the error shape, because the model needs to see it to recover.

Guardrails before autonomy

An agent without limits is an incident waiting to happen. Set a maximum number of steps, a timeout per tool call and a total budget for the run. Make destructive tools require confirmation or a dry-run flag. Decide whether the agent may call tools in parallel or strictly sequentially, and log every step with the model name, arguments and result. These controls belong in the builder so you can test them, and they belong again in your production service so the exported configuration is not the only place they exist.

Reading the trace to find the real bug

The trace separates three failure classes. Prompt failures show up as the agent skipping a tool or asking the user for information it already has. Schema failures appear as missing or invalid arguments. Reasoning failures show as correct calls chained in the wrong order. Fix them in that order: clarify the system prompt, then tighten the schema, then change the model tier. If a frontier model fixes it but the small model does not, that is a routing decision — send hard tasks to the frontier tier and routine tasks to a cheaper one through model routing.

Honest comparison

Capability	Plugsky agent builder	Framework-only	Hosted agent product
Visual tool definitions	Yes, JSON Schema with descriptions	Code only	Limited presets
Live model testing	30+ models from one catalogue	You configure each	Vendor models only
Trace inspection	Step-by-step tool and result view	Add logging yourself	Varies
Loop guardrails	Max steps, timeouts, stop conditions	You implement	Usually fixed
Export	Configuration you port into your service	Already code	Locked to platform
Deployment control	Plugsky cloud, VPC, on-prem, air-gapped	Wherever you run it	Vendor cloud only

Frequently asked questions

Do I need to write code to use the builder?

No. The builder is visual; you write prompts, tool descriptions and schemas, then export the configuration and port the loop into your own service for production.

Which models can I test with?

Any model in the live catalogue — 30+ models from free tiers to frontier reasoning — selectable per run.

Is function calling live on Plugsky?

Yes. Function calling, streaming, JSON mode, embeddings, RAG and agents are live; check the docs for the per-model capability matrix.

Can the agent remember context across runs?

Memory is available. Decide what persists — user preferences, prior tool results — and keep the rest scoped to a single run to avoid stale context.

How do I stop runaway loops?

Set maximum steps, per-tool timeouts and a total run budget, and make destructive tools require confirmation.

Can I use this with LangChain or AutoGen?

Yes. Treat the exported configuration as the design and implement the loop in your framework of choice; Plugsky's endpoint is OpenAI-compatible.

Does the builder run tools for me?

It invokes tools you define so you can inspect behaviour. Production tool execution should stay in your service, where credentials and side effects are controlled.

Is there a free way to try it?

Yes. The free plan includes 2 free AI models (plugsky-micro and plugsky-lite) with no card required.