← All workshopsRead Does Botie Need to Choose the Next Step?

Agents

Choose how much control Botie needs

A maintenance fault needs another lookup

The maintenance desk already has a working lookup workflow. Decide whether its tested results justify giving Botie more choice over the next operation.

BOTIE

The workflow now fails one of three required task cases. The agent passes all three within the same limits. What should the declared selection policy do?
The inputs we are comparing

Controllers tested on the same requests

The fixed workflow already does the job

[
  {
    "name": "rules",
    "rank": 0,
    "passed": [
      true,
      false,
      false
    ],
    "latency_ms": [
      10,
      12,
      11
    ],
    "cost_units": [
      0,
      0,
      0
    ]
  },
  {
    "name": "workflow",
    "rank": 1,
    "passed": [
      true,
      true,
      true
    ],
    "latency_ms": [
      600,
      800,
      700
    ],
    "cost_units": [
      2,
      2,
      2
    ]
  },
  {
    "name": "agent",
    "rank": 2,
    "passed": [
      true,
      true,
      true
    ],
    "latency_ms": [
      1200,
      1600,
      1400
    ],
    "cost_units": [
      5,
      6,
      5
    ]
  }
]

A fault needs a result-dependent lookup

[
  {
    "name": "rules",
    "rank": 0,
    "passed": [
      true,
      false,
      false
    ],
    "latency_ms": [
      10,
      12,
      11
    ],
    "cost_units": [
      0,
      0,
      0
    ]
  },
  {
    "name": "workflow",
    "rank": 1,
    "passed": [
      true,
      true,
      false
    ],
    "latency_ms": [
      600,
      800,
      700
    ],
    "cost_units": [
      2,
      2,
      2
    ]
  },
  {
    "name": "agent",
    "rank": 2,
    "passed": [
      true,
      true,
      true
    ],
    "latency_ms": [
      1200,
      1600,
      1400
    ],
    "cost_units": [
      5,
      6,
      5
    ]
  }
]

Task, time and cost requirements

Same in both cases

{
  "min_passes": 3,
  "max_latency_ms": 2000,
  "max_cost_units": 20
}
Result for the fixed workflow already does the job
workflow
Choose an explanation to test
A slower multi-agent candidate passes the same tasks

This check uses the pieces from “Choose a controller from task evidence”.

BOTIE

Workflow and agent both fail a required task. The multi-agent candidate passes, but one request takes longer than the declared latency limit. Predict the selection before running the existing construction.
Controllers tested on the same requests
[
  {
    "name": "workflow",
    "rank": 1,
    "passed": [
      true,
      false
    ],
    "latency_ms": [
      300,
      500
    ],
    "cost_units": [
      2,
      2
    ]
  },
  {
    "name": "agent",
    "rank": 2,
    "passed": [
      true,
      false
    ],
    "latency_ms": [
      900,
      1000
    ],
    "cost_units": [
      4,
      5
    ]
  },
  {
    "name": "multi-agent",
    "rank": 3,
    "passed": [
      true,
      true
    ],
    "latency_ms": [
      1200,
      2400
    ],
    "cost_units": [
      7,
      8
    ]
  }
]
Task, time and cost requirements
{
  "min_passes": 2,
  "max_latency_ms": 2000,
  "max_cost_units": 20
}

Use a number, text, or an array matching the requested result. For a matrix, put each row in brackets. Saving your prediction runs your current construction on these inputs and checks the earlier examples too.

Build a construction in the workspace above first. This case will test those pieces.

What evidence would make a multi-agent controller worth including in the next comparison, and what handoff failure would you test?

Choose the amount of autonomy from the task’s measured needs. Preserve the acceptance checks when changing the controller.

The small invented trial set is a teaching comparison. It omits repeated model runs, uncertainty intervals and production security review.

Work through the Python challenges →
Continue the courseGive Botie a Typed Job with Pydantic AI →
← What Should Botie Do After a Tool Replies?