Agent Layer

Agent Layer

This layer is SIGNET's distinctive contribution. No existing procurement standard represents synthetic agents as first-class, governed, accountable participants. The agent layer does — and it is always cleanly separable from third-party standards, so projecting a SIGNET document down to OCDS simply omits it (see Standards Mapping).

Object	Purpose
SyntheticAgent	An AI agent operating as a first-class Party.
AgentCapability	A declared, discoverable capability (A2A Agent Card aligned).
Mandate	The bounded authority granted to an agent — the structural guarantee it cannot exceed its remit.
Decision	The accountability record: what was decided, by whom, under what authority, from what, why.
Policy	Machine-readable, human-auditable rules — "rules as code".

How the agent layer enforces governance

The three guarantees that make a SIGNET network governed rather than merely automated:

1. Bounded authority. A Mandate lists exactly which capabilities an agent may exercise, the hard constraints it must respect, and the approvalThresholds above which a human must approve. An agent acting outside its mandate is, by construction, non-conforming.
2. Accountable action. Every material agent action produces a Decision recording the agent, the mandate relied on, the inputs considered, the policies applied, the rationale, the outcome, any human approval, and cryptographic provenance.
3. Auditable rules. Every Policy carries both an executable expression and a humanReadable statement of the same rule — so the rule that governs the agent is the rule a human can review.

See it run (v0.5.0). The repository ships a runnable proof of exactly these three guarantees: the agent/ demonstration (npm run agent). A synthetic agent awards a €12M contract — bounded by a Mandate whose €10M autonomous-value ceiling forces human approval, applying the published MAT evaluation Policy — and the runner then verifies the agent's output is conformance-clean: every Decision, Evaluation, and Award validates, the five-event hash chain holds, and tampering is detected. See Repository Structure → agent/.

---

SyntheticAgent

An AI agent operating within the network as a first-class Party (partyType: agent, role syntheticAgent).

Field	Type	Card.	Definition
id	Identifier	1	Agent DID.
name	string	1	Agent name.
operatedBy	Identifier	1	The Party accountable for the agent.
model	string	0..1	The underlying model family (model-agnostic; for provenance only).
capabilities	AgentCapability[]	1..*	Declared capabilities.
mandate	Identifier	1	The Mandate governing the agent.
agentCard	string (URI)	0..1	A2A Agent Card location (/.well-known/agent.json).

operatedBy answers the most important governance question — who is accountable for this agent? — and model is recorded for provenance, not for capability gating: the model is deliberately model-agnostic.

---

AgentCapability

A declared capability, aligned to the A2A Agent Card model so capabilities are discoverable across organisational boundaries.

Field	Type	Card.	Definition
name	string	1	Capability name, e.g. evaluate.submission, negotiate.price.
description	string	0..1	What it does.
inputSchema	object	0..1	Expected input.
outputSchema	object	0..1	Produced output.

---

Mandate

The authority granted to an agent — what it may do, within what limits, and where human approval is required. The structural guarantee that agents cannot exceed their remit.

Field	Type	Card.	Definition
id	Identifier	1	Mandate identifier.
agent	Identifier	1	The agent governed.
grantedBy	Identifier	1	The Party granting authority.
permittedCapabilities	string[]	1..*	Capabilities the agent may exercise.
constraints	Policy[]	0..*	Hard limits (e.g. max discount, value ceiling).
approvalThresholds	Policy[]	0..*	Conditions requiring human-in-the-loop approval.
scope	object	1	Data/entity scope the agent may operate within (sandbox boundary).
validity	Period	0..1	Time bound on the mandate.

The lifecycle of a mandate is itself event-sourced: mandate.granted and mandate.revoked are recorded as Events (see eventType codelist).

---

Decision

A record of a decision taken in the network, by a human or synthetic agent, with the inputs and rationale that produced it. The backbone of accountability.

Field	Type	Card.	Definition
id	Identifier	1	Decision identifier.
decisionType	string	1	See decisionType codelist: admissibility, evaluation, award, negotiationMove, qualification.
madeBy	Identifier	1	The agent (human or synthetic) that decided.
underMandate	Identifier	0..1	The mandate relied on, where the decider is synthetic.
inputs	Identifier[]	0..*	The objects considered (submissions, credentials, policies).
policiesApplied	Identifier[]	0..*	The policies applied.
rationale	string	1	Human-readable rationale.
outcome	object	1	The decision outcome.
humanApproval	Identifier	0..1	Where a mandate threshold required it, a reference to the human approval record. Under the identity profile this SHOULD resolve to a verifiable Approval (pseudonymous approver, role, delegation-of-authority credential, provenance).
provenance	Provenance	1	Full Provenance.

A Decision record satisfies the "material decision" record-keeping that regulation increasingly requires — including the assessment-summary and decision-record obligations of the UK Procurement Act 2023, and the documentation expectations for AI-assisted evaluation under the EU AI Act.

See the worked instance: examples/award-decision.json — a single object that records what was decided, which synthetic agent decided it, under which mandate, from which inputs, under which policy, with what rationale, with which human approval, and with cryptographic provenance.

---

Policy

A machine-readable, human-auditable rule. Policies express eligibility criteria, evaluation models, approval routing, agent constraints, and compliance rules. "Rules as code" made concrete.

Field	Type	Card.	Definition
id	Identifier	1	Policy identifier.
policyType	string	1	See policyType codelist: eligibility, evaluation, approval, constraint, compliance.
expressionLanguage	string	1	The language the rule is written in, e.g. rego, dmn, cel.
expression	string	1	The executable rule.
humanReadable	string	1	A plain-language statement of the same rule.
version	string	1	Policy version.
issuedBy	Identifier	1	The governing Party.

The dual requirement is mandatory: a Policy MUST carry both expression (machine- executable) and humanReadable (auditable), so the same rule governs agents and is reviewable by humans. Policies are referenced from the process layer as governingPolicies, eligibilityCriteria, and evaluationCriteria.

See the worked instance: examples/policy-evaluation.json — a Most-Advantageous-Tender scoring model expressed in Rego with a plain-language equivalent.

Where to go next

• Trust Layer — Provenance, Events, and Consent.
• Process Layer — where Decisions and Policies attach to the lifecycle.
• Worked Examples.