n8n AI Agent Not Calling Tools: tool_choice, Provider Compatibility, and Memory

n8n AI Agent Not Calling Tools: tool_choice, Provider Compatibility, and Memory

An n8n AI Agent can have a Code Tool, HTTP Request Tool, or MCP Tool attached, include a system instruction that says the tool must be used, and still return a direct answer. Adding more emphatic prompt language is not a reliable fix. First determine where the failure happens: the model never emitted a tool call, n8n could not parse the call, the tool ran but its result never reached the agent, or multi-turn memory dropped the execution messages.

The practical distinction is simple: if the tool node never starts in the execution graph, investigate tool selection and request shape; if the tool node starts and fails, investigate arguments, schema, credentials, and the downstream service. For the first case, n8n issue #31135 reports a concrete boundary: AI Agent v3 can send tools[] to an OpenAI-compatible provider on the first iteration without explicitly sending tool_choice: required. For a model with a weak tool-calling prior, the default auto behavior permits a normal text answer.

This article keeps official reports, local evidence, and unverified claims separate. The issue describes behavior observed in n8n. The local lab only proves that an absent tool_choice and required are different request contracts; it does not present simulated output as a benchmark of Mistral, Ollama, or vLLM.

Start with the execution graph, not the final answer

Open the execution and inspect the AI Agent, chat model, and tool nodes.

Evidence	Meaning	Next check
Tool node never ran	No executable tool call reached the tool	tool_choice, provider compatibility, model capability, tool descriptions
Tool node ran and schema validation failed	The model called the tool with invalid arguments	JSON Schema, required fields, types, nested data
Tool succeeded but agent reports failure	Result parsing or output shape failed	Tool response contract and error branches
Single turn works, later turns skip execution	Memory may omit tool calls or results	Stored message types and business state
One provider fails and another works	The compatible API may not fully implement tool calling	Raw request and response capture

The official n8n Tools Agent documentation states that the node uses LangChain’s tool-calling interface to describe available tools and schemas. Exposing a tool does not guarantee that the model selects it; selection still depends on the request contract, provider implementation, and model behavior. Tools AI Agent node

Why “you must call the tool” can still be ignored

A prompt influences model behavior. tool_choice constrains the API request. They operate at different layers.

A first-iteration request may look like this:

{
  "model": "mistralai/Mistral-Small-24B-Instruct",
  "messages": [
    {"role": "system", "content": "You MUST call get_value."},
    {"role": "user", "content": "What is the value?"}
  ],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_value",
        "parameters": {"type": "object", "properties": {}}
      }
    }
  ]
}

The tools array says which functions are available. Without an explicit tool_choice, many compatible APIs treat the request as auto, which allows ordinary prose. A larger model may infer the intended action from the prompt; a smaller model or compatibility layer may simply guess an answer.

The narrow fix proposed in issue #31135 is to force tool use on the first iteration and return to auto after a tool result. Keeping required on every iteration can create repeated calls and prevent the agent from producing a final answer.

The local lab verifies the contract, not model quality

The reproducible files are in:

experiments/n8n-ai-agent-tool-choice/

The lab uses a deterministic OpenAI-compatible provider fixture with one explicit rule:

• tools present without tool_choice: required returns prose;
• tools present with required returns a structured tool call.

Twenty repeated requests produced:

Scenario	Runs	Tool calls	Text responses
First iteration without tool_choice	20	0	20
First iteration with proxy-injected required	20	20	0

Those numbers are deterministic fixture results, not a real-model accuracy measurement. They prove only that a provider treating a missing value as auto has a different contract from one receiving required.

The lab also verifies three boundaries:

1. no injection when no tools are bound;
2. no override of explicit none or another non-default choice;
3. no forced choice after the history already contains an assistant tool call.

The core logic is intentionally small:

def should_force_tool_use(payload):
    tools = payload.get("tools") or []
    if not tools:
        return False

    for message in payload.get("messages") or []:
        if message.get("role") == "assistant" and (
            message.get("tool_calls") or message.get("function_call")
        ):
            return False

    return payload.get("tool_choice") in (None, "auto")

The directory contains request samples, a generated verification report, and five passing unit tests.

A temporary proxy without modifying n8n source

The preferred long-term outcome is an upstream fix verified against the exact n8n release in use. When a production workflow must keep using a specific compatible provider, a narrow proxy can sit between n8n and that provider:

n8n AI Agent
  -> OpenAI-compatible proxy
     -> inspect tools
     -> detect the first iteration
     -> inject tool_choice=required only there
  -> provider

A production proxy must:

• allow only controlled upstreams and models;
• avoid logging API keys or full sensitive prompts;
• enforce request-size, timeout, and concurrency limits;
• preserve explicit none, named choices, and later iterations;
• log whether it injected a choice, the model, tool count, and a redacted request digest;
• provide a fast rollback switch.

Do not blindly modify arbitrary JSON in a Code node and forward it. A proxy that cannot identify iteration boundaries may force the second iteration too, producing repeated emails, refunds, deployments, or an infinite tool loop.

When a tool call exists but execution still fails

At this point tool_choice is no longer the primary cause. Continue with four checks.

Does the provider support the actual schema?

Public issues include cases where a model emits a call and n8n reports Received tool input did not match expected schema. Common causes include:

• an object schema receiving a string;
• required fields that do not match the real tool input;
• nested objects or arrays flattened by a compatibility layer;
• enum values outside the declared set;
• different field names between the declared tool and a Tool Workflow.

Reduce the workflow to one parameter-free constant tool, then add arguments back one at a time. If the constant tool cannot run, investigate provider and request shape. If it runs but a complex schema fails, investigate the schema.

Are tool names and descriptions competing?

Issue #15883 describes identical prompts that sometimes call one tool, sometimes both, and sometimes none. Similar descriptions create an unstable routing problem.

Prefer action-and-object names:

get_customer_order_status
create_customer_refund_draft
send_refund_confirmation_email

Avoid vague siblings such as customer_tool, support_tool, and order_helper. State the input, output, side effects, and prohibition conditions in every description.

Does “OpenAI compatible” include full tool calling?

A compatible endpoint often guarantees only a similar URL and a subset of fields. It may not guarantee:

• every tool_choice variant;
• parallel calls;
• streamed argument deltas;
• the same JSON Schema support;
• stable call IDs and tool-result linkage;
• complete multi-turn message preservation.

Capture the raw request and response. Pin the model and provider version. Keep one minimal known-good sample.

Is the tool result structured for the agent?

Long prose, HTML, or a huge JSON object can hide the load-bearing value. Prefer a stable result contract:

{
  "ok": true,
  "data": {"value": 42},
  "error": null,
  "audit_id": "tool-20260726-001"
}

The model can read ok and data, while the execution and downstream systems retain a durable audit identifier.

Why multi-turn agents can claim execution without executing

Issue #14361 reports memory backends that persist user input and final assistant output but omit the tool call and tool result. The model later sees a pattern where a user requested an action, the assistant text said it succeeded, and the conversation continued. It may reproduce the success text without performing the action.

Debug this by:

1. inspecting the exact message types persisted by the memory backend;
2. verifying that call IDs, argument digests, and tool results are stored;
3. never treating the assistant’s claim as authoritative business state;
4. storing order IDs, approval IDs, and task status in a database rather than only in conversation memory;
5. returning an audit ID from every side-effecting tool and re-querying authoritative state when needed.

Memory helps the model understand a conversation. It must not be the only ledger for payments, orders, email delivery, or deployments.

A practical debugging order

1. Check whether the tool node starts
2. Keep one parameter-free constant tool
3. Repeat one fixed input 10 times
4. Capture the provider request and response
5. Inspect tools and tool_choice
6. Compare with a model known to support tool calling
7. Reintroduce schema, multiple tools, and memory gradually
8. Verify tool results and cross-turn state
9. Add audit IDs and idempotency keys for side effects

Change one variable at a time. If the model, prompt, schema, provider, and memory are all changed together, a successful run still does not identify the actual fix.

The reproducible lab is available in the public GitHub directory. It requires no API key and reruns both the request-contract fixture and five unit tests.

Final decision

“n8n AI Agent not calling tools” is a symptom shared by several failure classes: a missing first-iteration tool_choice, incomplete provider compatibility, weak tool selection, invalid schema, ambiguous descriptions, result parsing, and missing tool messages in memory.

The load-bearing evidence is not a sentence saying “I called the tool.” It is an execution that contains a real tool call, a tool node result, persisted state, and an audit record for external side effects. Prompts can influence the model; they cannot replace the request contract or execution evidence.

Continue with n8n AI Workflow error handling, retries, and cost monitoring, AI Agent Tool Use: registry, validation, and audit, and self-hosted n8n AI workflows.

Topic path / AI workflows

Continue through the production automation path

The workflow hub connects self-hosting, queue mode, webhooks, retries, observability and n8n implementation cases into one production-oriented learning path.

Open workflow hub → Review webhook hardening Subscribe for updates