AI agents API — build agents that call your APIs

An AI agent is an LLM that decides which tools to call, in what order, with what arguments, until the user's goal is met. Plugsky's agent API gives you the four primitives every production agent needs: tools, function calling, memory, and a planning loop.

What is the AI agents API?

Plugsky exposes agent capabilities through three layers:

1. Function calling on every chat model — same shape as OpenAI tools array.
2. Agent runs endpoint — POST a goal + tools, get back a full run with steps, tool calls, and final output.
3. Agent Cloud — hosted agents with memory, schedule, webhooks, and observability built in.

Tools and function calling

tools = [
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "Get current weather for a city",
            "parameters": {
                "type": "object",
                "properties": {
                    "city": {"type": "string"}
                },
                "required": ["city"]
            }
        }
    }
]

resp = client.chat.completions.create(
    model="plugsky-pro",
    messages=[{"role":"user","content":"What's the weather in Dubai?"}],
    tools=tools,
)
if resp.choices[0].message.tool_calls:
    call = resp.choices[0].message.tool_calls[0]
    print(call.function.name, call.function.arguments)

Plugsky's tools array matches the OpenAI shape exactly, so any OpenAI-compatible agent framework (LangChain, LlamaIndex, Vercel AI SDK, smolagents) works out of the box.

Memory and context

Plugsky Agent Cloud persists:

• Conversation history per agent / per user
• Tool call logs with full request / response payloads for replay
• Vector memory — long-term recall backed by Plugsky's RAG API
• Structured memory — typed key-value store for agent state

Memory is namespaced per workspace and per agent, with row-level encryption at rest.

Agent orchestration

For agents that need a planning loop, retry logic, or multi-step reasoning, use the Agent Runs endpoint:

curl -X POST https://api.plugsky.com/v1/agents/acme-sales/agent/runs \
  -H "Authorization: Bearer sk-live-..." \
  -H "Content-Type: application/json" \
  -d '{
    "input": "Find the Q3 renewal for Acme Corp and draft a follow-up email",
    "tools": ["crm.lookup", "gmail.draft", "calendar.find"],
    "max_steps": 10
  }'

You get back a full run with each step's reasoning, tool call, and result — perfect for debugging and audit.

Production agent patterns

• Tool-use agents: search, retrieval, code execution, file ops
• Multi-agent: orchestrator + specialist sub-agents
• Human-in-the-loop: pause for approval before sensitive tool calls
• Scheduled: cron-triggered agents that produce a daily report
• Webhook-triggered: agents that respond to GitHub, Stripe, or HubSpot events

Frequently asked questions