Strategy Blueprint

A strategy blueprint declares a parameterized evaluation strategy as a YAML file. It defines condition-based entry and exit rules, signal dependencies, evaluation criteria, and risk control thresholds. Strategy blueprints are consumed by domain-specific modules that build on the platform SDK.

Blueprints are validated at load time by the StrategyBlueprint Pydantic schema in agent_core.blueprints.strategy. Invalid blueprints fail loudly — there are no silent defaults for required fields.

Strategy blueprint files live under blueprints/strategies/*.yaml in a domain repo.

Top-Level Identity Fields

id, name, and version are required. All other top-level fields default to empty or None.

id: confidence-threshold           # Unique strategy identifier. Kebab-case by convention.
name: Confidence Threshold         # Human-readable display name. Required.
version: "1.0.0"                   # Semantic version string. Required.
description: |                     # Optional description.
  Activates when the primary confidence signal crosses above threshold
  and data quality conditions are confirmed.

required_agents:, required_mcps:, required_signals:

Declares dependencies. All three are lists of strings and default to empty lists.

required_agents:                   # Agent blueprint IDs that must produce input signals.
  - data-collector
  - analyzer

required_mcps:                     # MCP server names needed by this strategy.
  - scoring-mcp

required_signals:                  # Named signals expected from required agents.
  - confidence_score
  - data_quality_score
  - accuracy_metric

parameters: Block

Declares named strategy parameters with type constraints. Each parameter requires name and type. Optional fields: default, description, min_value, max_value.

Supported type values: int, float, str, bool, list. Common aliases are accepted: string → str, integer → int, number → float, boolean → bool, array → list.

parameters:
  - name: threshold
    type: float
    default: 0.7
    description: Minimum confidence score to activate.
    min_value: 0.0
    max_value: 1.0

  - name: lookback_window
    type: int
    default: 20
    description: Number of periods to consider.
    min_value: 5
    max_value: 500

  - name: mode
    type: str
    default: "standard"

  - name: enabled
    type: bool
    default: true

  - name: target_metrics
    type: list
    default: ["accuracy", "precision"]

entry_conditions: Block

Conditions that must be met to activate the strategy. Uses a ConditionGroupConfig with a logic operator (and or or, case-insensitive) and a non-empty conditions list.

Each condition uses field, operator, and value. Supported operators: eq, neq, gt, gte, lt, lte, in, not_in, between. A condition may alternatively use type for structured domain-specific conditions (e.g. type: threshold_breach) instead of the field/operator/value pattern.

entry_conditions:
  logic: and                       # and | or (default: and). Accepts AND/OR.
  conditions:
    - field: confidence_score
      operator: gte
      value: 0.8

    - field: data_quality_score
      operator: gte
      value: 0.7

    - field: status
      operator: in
      value: ["ready", "verified"]

exit_conditions: Block

Conditions that trigger strategy deactivation. Same structure as entry_conditions.

exit_conditions:
  logic: or
  conditions:
    - field: confidence_score
      operator: lt
      value: 0.3

    - field: error_rate
      operator: gt
      value: 0.05

    - field: accuracy_metric
      operator: between
      value: [0.0, 0.4]

evaluation: Block

Configures how this strategy is measured. primary_metric is required — all other fields are optional.

evaluation:
  primary_metric: accuracy         # Required. Primary performance metric for ranking.
  metrics:                         # All metrics to compute.
    - accuracy
    - precision
    - recall
    - f1_score
    - latency_p95
  benchmark: baseline-strategy     # Strategy ID for comparative benchmarking.
  lookback_window: 100             # Periods to include. Must be > 0.
  min_activations_threshold: 10    # Minimum activations for statistical significance. Must be > 0.
  persistence:                     # Optional. Persist evaluation scores to DynamoDB.
    enabled: true
    table_env: EVALUATION_TABLE    # Env var holding the DynamoDB table name.
    retention_days: 90

persistence stores evaluation scores independently in DynamoDB, regardless of the evaluation provider configured on agent blueprints. It is an EvaluationPersistenceConfig and defaults to None (disabled).

risk_controls: Block

Hard limits enforced independently of evaluation logic. All fields are optional; rate values must be between 0.0 and 1.0.

risk_controls:
  max_daily_error_rate: 0.03       # Halt if daily error rate exceeds 3%.
  max_degradation_halt: 0.10       # Halt if quality degradation from baseline exceeds 10%.
  circuit_breaker:                 # Arbitrary dict for domain-specific circuit-breaker config.
    consecutive_failures: 5
    pause_periods: 10

execution_modes: Block

Controls in which execution environments this strategy is active.

execution_modes:
  simulation: true                 # Active in simulation / testing environment.
  staging: true                    # Active in staging.
  production: false                # Disabled in production until promoted.

tags: Block

tags on StrategyBlueprint is a dict[str, str] — key-value pairs, not a list.

tags:
  team: data-science
  category: classification
  priority: high

Complete Example 

id: confidence-threshold
name: Confidence Threshold Strategy
version: "2.0.0"
description: |
  Activates when confidence score crosses above threshold with data quality
  confirmation. Exits on error threshold breach or signal reversal.

required_agents:
  - data-collector
  - analyzer

required_mcps:
  - scoring-mcp

required_signals:
  - confidence_score
  - data_quality_score
  - accuracy_metric

parameters:
  - name: threshold
    type: float
    default: 0.7
    description: Minimum confidence score to activate.
    min_value: 0.0
    max_value: 1.0
  - name: lookback_window
    type: int
    default: 20
    min_value: 5
    max_value: 500

entry_conditions:
  logic: and
  conditions:
    - field: confidence_score
      operator: gte
      value: 0.8
    - field: data_quality_score
      operator: gte
      value: 0.7

exit_conditions:
  logic: or
  conditions:
    - field: confidence_score
      operator: lt
      value: 0.3
    - field: error_rate
      operator: gt
      value: 0.05

evaluation:
  primary_metric: accuracy
  metrics:
    - accuracy
    - precision
    - recall
    - latency_p95
  benchmark: baseline-strategy
  lookback_window: 100
  min_activations_threshold: 15
  persistence:
    enabled: true
    table_env: EVALUATION_TABLE
    retention_days: 90

risk_controls:
  max_daily_error_rate: 0.025
  max_degradation_halt: 0.08
  circuit_breaker:
    consecutive_failures: 4
    pause_periods: 8

execution_modes:
  simulation: true
  staging: false
  production: false

tags:
  team: data-science
  category: classification

Schema Reference

Field Type Required Default Description
id str Yes — Unique strategy identifier (kebab-case recommended)
name str Yes — Human-readable display name
version str Yes — Semantic version string
description str No "" Strategy description
required_agents list[str] No [] Agent blueprint IDs that must produce input signals
required_mcps list[str] No [] MCP server names needed
required_signals list[str] No [] Named signals expected from agents
parameters list[ParameterConfig] No [] Parameterized configuration with type constraints
entry_conditions ConditionGroupConfig No null Activation conditions
exit_conditions ConditionGroupConfig No null Deactivation conditions
evaluation StrategyEvaluationConfig No null Evaluation config (primary_metric required if set)
risk_controls RiskControlConfig No null Hard limit guardrails
execution_modes ExecutionModes No simulation=true Environment gates
tags dict[str, str] No {} Arbitrary key-value metadata tags