Portable Agent Intelligence: AIR Quickstart

The Agent Intelligence Record (AIR) packages everything an agent has learned into a portable, verifiable bundle that travels with the agent. When an agent moves to a new platform or environment, AIR ensures it arrives knowing what it already knows.

AIR bundles are organized into three layers: what governance decided, what your organization defined, and what the agent learned through experience. Each layer is independent — export one, two, or all three depending on what matters for your use case.


Before you start

  • agentpk 0.3.3 or later: pip install --upgrade agentpk
  • An agent packaged with agentpk
  • Optionally: a Nomotic account for full intelligence export

Step 1 — Pack with memory

Add --memory to any agentpk pack command:

agentpk pack ./my-agent --memory

Pack with code analysis and memory together:

agentpk pack ./my-agent --analyze --memory

Step 2 — Choose your components

By default, --memory includes all components your environment can produce. To include specific ones:

agentpk pack ./my-agent --memory --memory-components fingerprint,trust,org_context

Layer 1 — Governance Record

Produced automatically from your agent’s governance history.

ComponentFileWhat it contains
auditaudit.jsonlHash-chained transaction history, tamper-evident
fingerprintfingerprint.jsonStatistical behavioral profile — action distributions, target patterns, outcome rates
trusttrust.jsonTime-series of trust score changes with labeled events

Layer 2 — Institutional Context

Configured in Nomotic’s Context and Governance modules.

ComponentFileWhat it contains
org_contextorg_context.jsonVocabulary, values, domain anchors, and escalation patterns
compliance_statecompliance_state.jsonActive regulatory frameworks (HIPAA, PCI-DSS, SOC2), policy versions, and data classifications in effect at export time

Layer 3 — Operational Intelligence

What the agent learned through experience — not governance rules, but judgment developed by doing the job.

ComponentFileWhat it contains
domain_modeldomain_model.jsonLearned entities, relationships, edge cases encountered, and thresholds calibrated through experience
interaction_patternsinteraction_patterns.jsonRequest patterns, ambiguity resolutions, communication preferences, and system integration quirks
knowledge_stateknowledge_state.jsonDistilled insights with confidence scores and source tagging (governance-derived vs. experience-derived)

Step 3 — Inspect the bundle

agentpk inspect my-agent-1.0.0.agent

The AIR section shows all three layers and their components:

-- Agent Intelligence Record (AIR) ──────────────────────────────────
Version          1.1
Issuing platform nomotic 0.9.0
Exported         2026-04-12T00:00:00Z
Redaction        standard

Layer 1  Governance Record
         fingerprint · trust · audit

Layer 2  Institutional Context
         org_context · compliance_state

Layer 3  Operational Intelligence
         domain_model · interaction_patterns · knowledge_state

License          proprietary (re-export prohibited)
Hashes           ✓ all verified

Step 4 — Verify the bundle

agentpk validate my-agent-1.0.0.agent

Validation recomputes SHA-256 for every component and compares against air.json. Any mismatch rejects the package.


What’s inside the archive

my-agent-1.0.0.agent
  manifest.yaml
  checksums.sha256
  src/agent.py
  requirements.txt
  intelligence/
    air.json                    ← bundle manifest (always present)
    fingerprint.json            ← Layer 1
    trust.json                  ← Layer 1
    audit.jsonl                 ← Layer 1 (if included)
    org_context.json            ← Layer 2
    compliance_state.json       ← Layer 2 (if included)
    domain_model.json           ← Layer 3 (if included)
    interaction_patterns.json   ← Layer 3 (if included)
    knowledge_state.json        ← Layer 3 (if included)

Full intelligence export with Nomotic

Without a Nomotic account, --memory embeds a spec-compliant stub recording which components were requested. No behavioral data.

With a Nomotic account, the full export hydrates each component from your agent’s live data — the actual fingerprint built from its transaction history, real trust trajectory, compliance frameworks you configured, and operational intelligence the agent developed.

export NOMOTIC_API_KEY="nm_live_..."
agentpk pack ./my-agent --memory

If the API key is not set, agentpk falls back to the stub and prints an advisory. The package is valid either way.


Redaction profiles

ProfileWhat it keepsUse when
minimalRemoves free-text onlyInternal transfers
standardRemoves PII identifiers, keeps statistical shapeDefault
strictVerdict, timestamp, and hash per audit record onlyHIPAA / PHI environments

The redaction profile is recorded in air.json. The audit hash chain tail is always present regardless of profile.


Importing an AIR bundle

When an agent with an AIR bundle is imported into a Nomotic environment, the runtime rehydrates using this fallback hierarchy:

  1. Full rehydration — all components verified: fingerprint seeds behavioral consistency, trust trajectory sets starting score, compliance frameworks apply immediately
  2. Governance warm-start — Layer 1 only: fingerprint and trust rehydrated, institutional context resets to defaults
  3. Context cold-start — Layer 2 only: trust and fingerprint reset to archetype defaults, org vocabulary applies
  4. Bare cold-start — no components verified: treated as new agent, warning logged listing failed components

The import event is recorded as "event": "air_import" in the trust trajectory so post-import behavior is distinguishable from pre-import history.


Licensing your intelligence

"intelligence_license": {
  "default": "proprietary",
  "components": {
    "fingerprint": "cc-by-nc-4.0"
  },
  "permitted_uses": ["rehydration", "governance"],
  "prohibited_uses": ["fine-tuning", "benchmarking", "re-export"]
}

Per-component overrides are supported — share behavioral statistics while keeping audit records and operational intelligence proprietary.


Python SDK

from agentpk import pack

result = pack(
    "./my-agent",
    analyze=True,
    memory=True,
    memory_components=["fingerprint", "trust", "org_context", "compliance_state"],
)

print(result.package_path)       # PosixPath('./my-agent-1.0.0.agent')
print(result.air_components)     # ['compliance_state', 'fingerprint', 'org_context', 'trust']
print(result.trust_score)        # 87

The AIR open standard

AIR is an open specification published under CC BY 4.0. Any platform can produce or consume an AIR bundle without requiring Nomotic.

Schema files for all 8 components are available at agentpk.io/specs/air/v1.1.


Next steps

  • Configure organizational vocabulary in Context to populate org_context before export
  • Set active compliance frameworks in Governance → Compliance to populate compliance_state
  • Review your agent’s trust trajectory in Governance → Trust before exporting
  • See the AIR Schema Reference for complete field documentation


Was this article helpful?

On this page