One post tagged with "tool-calling"

We are introducing the AI Orchestrator — a new worker that runs an AI agent inside your workflow. Instead of wiring every step by hand, you select the workers the agent may use and describe the task. The orchestrator chooses the parameters, calls the workers, adapts when data is missing, and writes the final answer.

It is the difference between a fixed pipeline and a delegation.

Why we built it​

Financial workflows rarely follow the same path twice. A market brief may need price data, news, and the economic calendar — but sometimes the news feed is empty, the calendar needs a wider window, or a symbol has no recent candles. A fixed workflow has to anticipate every branch. The AI Orchestrator can decide at runtime.

Common examples:

• prepare a morning market brief from several sources,
• compare price action with news and macro events,
• monitor a setup and alert only when a score is high enough,
• research a symbol with multiple data providers,
• remember preferences and lessons between scheduled runs.

How it works​

The orchestrator is a ReAct-style tool-calling agent:

1. It receives your prompt, the selected workers, and the run context.
2. Each selected worker is exposed as a tool with its parameter schema.
3. The model reasons, calls a worker, observes the result, and decides the next step.
4. It can retry with different parameters or choose another worker.
5. When it has enough information, it writes the final answer.

Every tool call is a real worker execution. That means API credits, rate limits, credentials, and side effects apply exactly as they would for a normal node. The trace is returned in the calls field so you can audit what happened.

What you get​

The node returns more than a text answer:

The output panel shows answer as its own group next to the calls and workers_used tables, so you can inspect both the result and the path that produced it.

Quick start​

1. Add the node​

Place AI Orchestrator after a trigger and before whatever consumes the result:

[Schedule] -> [AI Orchestrator] -> [Telegram]

2. Select the workers it may use​

Keep the set small and focused. For a market brief:

fetch_price
fetch_news
browse_calendar

Triggers and visualization/widget nodes are hidden automatically — they cannot be executed as tools. The orchestrator itself and parameter_loop are also excluded.

3. Write the prompt​

State the goal, output format, time range, and what to do when data is missing:

Prepare a concise EURUSD market brief for the London session.

Use the last 24 hours of price action, the most relevant news, and the
high-impact events in the next 48 hours. Compare the technical picture with
the news flow and call out disagreements.

Return three bullet points of evidence, a directional bias, and a confidence
score from 1 to 5 on a final line exactly like: SCORE: 4

If news is unavailable, say so explicitly and continue with price and calendar
data. Do not invent headlines.

4. Run it​

Run the workflow from the editor or wait for the trigger. Open the calls trace to see which workers were used and with which parameters.

Memory that lasts between runs​

Set memoryScope to control persistence:

• run — memory only inside the current run (default),
• workflow — session history and long-term facts persist for this workflow/node/owner,
• none — no memory is loaded or saved.

With workflow scope, the agent can use remember and recall tools. Save stable knowledge, not market snapshots:

• good: "Use the London timezone and report prices in EUR for this workflow."
• good: "The CoinGlass funding-rate endpoint is the preferred BTC derivatives source."
• not useful: "BTC is 86,500 right now."

Guardrails for production​

Long-running agents need limits. Open advancedSettings to tune:

• maxSteps — maximum reasoning/tool steps (default 15),
• maxRuntimeSeconds — wall-clock budget for the node (default 180),
• toolTimeout — maximum time for one delegated worker,
• tokenBudget — total token budget, 0 for unlimited,
• maxToolCalls — hard cap on worker invocations, 0 for unlimited,
• planningMode — react to act immediately, plan to write a plan first.

Identical worker calls with identical parameters are cached within a run. A retry with different parameters is a new call and counts against maxToolCalls.

Conditional alerts with a score​

The orchestrator does not decide which downstream nodes run. For deterministic branching, ask for a score and use a Condition node.

Prompt ending:

End with a line exactly like: SCORE: 4

Condition expression:

vars['Orchestrator']['score'] >= 4

Then send Telegram only on the true branch. You get the agent's reasoning plus a hard threshold you can audit.

Choosing between the orchestrator and a fixed pipeline​

Use the AI Orchestrator when the sequence of steps depends on what the data says. Use ordinary workers or Run Workflow when the sequence is known in advance.

A useful rule:

• Fixed steps -> normal workflow nodes.
• Deterministic workflow-to-workflow orchestration -> Run Workflow.
• Adaptive research across several workers -> AI Orchestrator.
• Trading analysis with structured orders -> AI Makler.

Learn more​

The full worker reference — parameters, output fields, memory details, examples, and troubleshooting — is available in the AI Orchestrator documentation.

The deep dive article covers the agent loop, tool selection, prompt engineering, and production patterns in more detail.

Availability​

AI Orchestrator is available now in the public worker library under the AI category. Add it to a workflow, select a few workers, and give it a task.

As always, validate automated outputs before acting on them. The orchestrator is a reasoning layer, not a guarantee of accuracy, and it should sit behind your risk controls.

This post is for educational purposes only and does not constitute investment advice. Financial markets involve risk. Always test workflows and use appropriate risk management.