How data passes between agents in a flow

When Agent A finishes, its output becomes available as input to Agent B. MilkyWay matches fields automatically where names and types match. You fill in the gaps.

---

The concept

Every agent declares an output_schema (what it returns) and an input_schema (what it accepts). When you draw a connection between two agents in the builder, MilkyWay reads both schemas and tells you:

• Which fields matched automatically
• Which fields need manual input from you
• Which mappings have type mismatches

Agent A — output

Agent B — input

price: number

✓ auto-matched

price: number

asset: string

✓ auto-matched

asset: string

timestamp: number

ignored

—

—

⚠ fill manually

action: stringfill in builder

MilkyWay matches fields by name and type automatically. Gaps appear as input fields in the visual builder's right panel.

---

Auto-matching example

Agent A (price_monitor) output schema:

{
  "price": { "type": "number" },
  "asset": { "type": "string" },
  "timestamp": { "type": "number" }
}

Agent B (risk_analyzer) input schema:

{
  "price": { "type": "number", "required": true },
  "asset": { "type": "string", "required": true },
  "action": { "type": "string", "required": true },
  "threshold": { "type": "number", "required": false }
}

What the builder shows when you connect them:

Automatic matches:
  price (number)  →  price (number)   ✓
  asset (string)  →  asset (string)   ✓

Manual input required:
  action    [__________]   required
  threshold [__________]   optional

price and asset match by name and type — they pass through automatically. action has no source in Agent A's output, so you type a static value ("sell", "hold", etc.).

---

Type coercion in mappings

MilkyWay performs safe coercions automatically. Unsafe coercions are flagged.

Safe (shown in green):

Source type	Target type	Notes
number	string	42 → "42"
boolean	string	true → "true"
number	boolean	0 → false, non-zero → true

Unsafe (shown with warning):

Source type	Target type	What happens
string	number	Warning — may fail if string isn't numeric
string	object	Warning — pass-through attempted
object	string	Warning — JSON-stringified

---

Manually mapping a field from a different name

If Agent A outputs count (number) and Agent B expects items (string):

1. Click the connection arrow on the canvas
2. Right panel shows Field Mappings
3. Click Add mapping
4. Select source field: count
5. Select target field: items
6. MilkyWay applies the number → string coercion automatically

The mapping appears:

count (number)  →  items (string)  [safe coercion]  ✓

---

When fields don't match

Agent A outputs { result: string }. Agent B expects { data: object }.

The builder shows:

⚠ Incompatible mapping:
  result (string)  →  data (object)
  Unsafe coercion. Use a transform agent between them,
  or choose a different Agent B.

Options:

1. Use a different Agent B that accepts string input
2. Insert a transform agent between A and B that converts the string to an object
3. Provide data as a static JSON value manually

---

Passing complex data (objects and arrays)

Objects and arrays pass through unchanged when the types match:

Agent A output:

{
  "analysis": {
    "risk_level": "high",
    "positions": [...]
  }
}

Agent B input schema:

{
  "analysis": { "type": "object" }
}

The entire analysis object passes through as-is. Agent B receives it exactly as Agent A returned it.

---

Debugging mapping issues

When a run fails with input validation error, enable development logging on your agent and look at what it received:

NODE_ENV=development npm run dev

Your agent will log the raw input before passing it to the handler:

→ execute  job=abc-123
  input received: {
    "price": 2847.32,
    "asset": "ETH",
    "action": "sell",
    "threshold": null    ← optional field not provided, came through as null
  }

Compare this to what the previous agent returned — the discrepancy tells you exactly which mapping is wrong.

---

Viewing the output at each step

From your flow status page at usemilkyway.com/flows/<flowId>, click any completed agent in the pipeline to expand its output:

[Price Monitor]  ✓ done  142ms
▾ View output
  {
    "price": 2847.32,
    "asset": "ETH",
    "timestamp": 1748995200
  }

[Risk Analyzer]  ✓ done  831ms
▾ View output
  {
    "risk_level": "safe",
    "recommendation": "Hold — price is stable."
  }

---

What's next

• Build a DeFi safety flow — field mapping in a real three-agent flow
• Schedule a recurring flow — run your flow on a schedule
• Chain agents without the builder — manual field mapping in code