KCC 05

Agent Contracts

The Agent Contract is the interface that makes agentic work governable across identity, inputs, outputs, tools, cost, confidence, HITL, observability, and decision trace.

Agent ContractAgent ContractNine surfacesKernelSurfaces 1-9
Created 2026-06-08 · v0.4.0

What a Surface Is

Every agent in a KCC-conformant organization implements nine surfaces. The first seven are operational; the last two are cognitive. A surface is a structured field in the agent's contract that the kernel validates before the agent is registered, that meta-agents (Butler, Token Guard) consult at invocation time, and that the Inspector Pipeline reads to detect patterns.

Surfaces are declared once per agent version. The kernel validates surface declarations before registration and rejects agents with missing or malformed surfaces.

Surfaces 8 and 9 are the most distinctive technical contribution of KCC. Most agent frameworks specify some form of Surfaces 1-7. Few specify confidence and decision trace as structured first-class fields enforced by the kernel.

The Nine Surfaces

#SurfaceLayerPurpose
1IdentityOperationalUniquely identify the agent and its provenance
2Input SchemaOperationalDeclare what the agent accepts as input
3Output SchemaOperationalDeclare what the agent produces
4Declared ToolsOperationalExhaustive tool declaration; lethal trifecta detection
5Cost EnvelopeOperationalResource budgets and ceilings
6HITL/HOTLOperationalTrust mode declaration
7ObservabilityOperationalTelemetry commitment
8ConfidenceCognitiveComputed confidence in output
9Decision TraceCognitiveStructured record of how the output was produced
Agent Contract Surfaces
1 Identity
2 Input
3 Output
4 Tools
5 Cost
6 HITL
7 Observability
8 Confidence
9 Decision trace

Why Nine Surfaces

The number is not arbitrary. The first seven address operational concerns that every agent must declare to operate within the KCC contract: how to identify the agent (1), what it accepts (2) and produces (3), what tools it uses (4), what it costs (5), how much human oversight it requires (6), and what telemetry it emits (7). The last two address cognitive concerns that distinguish KCC from less mature frameworks: how confident it is in its output (8), and how it arrived at that output (9).

The cognitive layer is what makes the system learnable. Without confidence and decision trace, the system can run agents but cannot learn from them. The Inspector Pipeline depends entirely on Surface 9. See Identity Surface, Input and Output Surfaces, Tools and Cost Surfaces, HITL and Observability, and Confidence and Decision Trace for the individual surface specifications.

Diagram

Surface Interactions

The surfaces are not independent. Several interactions are structurally important: Declared Tools (4) determines HITL/HOTL (6) because lethal trifecta detection forces HITL; Cost Envelope (5) constrains Observability (7) because telemetry that exceeds the envelope is itself a problem; Confidence (8) influences HITL/HOTL (6) because low confidence triggers HITL escalation; and Decision Trace (9) reflects all other surfaces. The kernel enforces these interactions. An agent declaration that is internally inconsistent (e.g., declared HOTL but flagged trifecta in tools) is rejected at registration.

A Complete Agent Declaration

A complete agent declaration includes all nine surfaces. Surfaces 8 and 9 are populated per-invocation in the output, but the schemas referenced are part of the agent's declaration. Here is the skeleton:

# Agent: <name>
# Version: <semver>
# Maturity: <L1|L2|L3>

identity:                  # Surface 1
  # ...

input_schema:              # Surface 2
  # ...

output_schema:             # Surface 3
  # ...

declared_tools:            # Surface 4
  # ...

cost_envelope:             # Surface 5
  # ...

hitl_hotl:                 # Surface 6
  # ...

observability:             # Surface 7
  # ...

# Surfaces 8 and 9 are populated per-invocation in the output,
# but the schemas referenced are part of the agent's declaration.