Skip to main content
The harness makes one of four decisions at every step. Actions are computed in both observe and enforce modes, but only applied in enforce mode.

​
Actions

​
allow

Proceed normally. No constraints are violated. 
Step 1: allow — budget ok, model compliant
This is the most common action. It means all hard caps (budget, tool calls, latency, energy) are within limits and compliance is satisfied.

​
switch_model

Route to a different model. Triggered when:
  • The current model is not in the compliance allowlist
  • KPI scoring indicates a better model choice
  • Budget pressure suggests a cheaper alternative
Step 3: switch_model — compliance violation, switching to gpt-4o-mini (gdpr allowlist)
In enforce mode, the harness substitutes the model. In observe mode, the original model is used and the trace records what would have happened.

​
deny_tool

Block a tool/function call. Triggered when max_tool_calls is reached. 
Step 5: deny_tool — tool call cap reached (10/10)
In enforce mode, the tool call is blocked. The agent receives a signal that the tool was denied.

​
stop

Halt agent execution. Triggered when:
  • Budget is exceeded
  • Latency cap is exceeded
  • Energy cap is exceeded
Step 7: stop — budget exceeded ($0.52 > $0.50 cap)
In enforce mode, the agent loop is stopped. In observe mode, execution continues and the trace records the violation.

​
Decision Priority

When multiple constraints are violated simultaneously, the harness applies this priority:
  1. Compliance — check first (switch_model or stop)
  2. Budget — check second (stop)
  3. Tool calls — check third (deny_tool)
  4. Latency — check fourth (stop)
  5. Energy — check fifth (stop)
  6. KPI scoring — soft optimization (switch_model or allow)

​
Hard vs Soft Controls

Hard controls trigger stop or deny_tool when limits are exceeded:
  • budget — max USD
  • max_tool_calls — max tool/function calls
  • max_latency_ms — max wall-clock ms per call
  • max_energy — max energy units
  • compliance — model allowlist
Soft controls influence model selection through KPI weights but never block execution:
  • kpi_weights — relative importance of quality, cost, latency, energy
  • kpi_targets — target values for KPI dimensions

​
Example: Combined Constraints

import cascadeflow

cascadeflow.init(mode="enforce")

with cascadeflow.run(
    budget=1.00,
    max_tool_calls=5,
    compliance="gdpr",
    kpi_weights={"quality": 0.6, "cost": 0.4}
) as session:
    result = await agent.run("Process EU customer data")

    for record in session.trace():
        print(f"Step {record['step']}: {record['action']} — {record['reason']}")
Examples on GitHub: examples/enforcement/basic_enforcement.py | examples/agentic_multi_agent.py