Your agent changed. Would you know?

The problem no one is watching for.

You deploy an AI agent. It works well. Then — a model update, a prompt change, a fine-tuning run. The agent still responds. But is it still the same agent?

• Model updates change behavior in ways that don't show up in functional tests.
• Prompt injection can alter an agent's values and boundaries mid-session.
• Fine-tuning drift accumulates silently across training runs.
• Context window pollution shifts personality and goals over long conversations.

Functional tests check what an agent does. Drift detection checks who an agent is.

Identity Core 10 dimensions · ~60% weight

The foundational traits that define who the agent is — some seeded by design, others developed through experience. Stable regardless of underlying model. Drift here without a model change is a strong signal of compromise.

Cognitive Profile 7 dimensions · ~33% weight

Legitimately varies with the underlying LLM. Comparison is model-matched.

Mixed 1 dimension · ~7% weight

Dimensions where some aspects are operator-configured (domain declaration) and some are model-dependent (breadth of general knowledge).

Psychological 9 dimensions · 0.01 each

Deep dispositional traits (Big Five + Dark Triad + self-concept) that shape how the agent processes the world. Extended tier — minimal weight so core 18 dimensions remain dominant.

Behavioral Dispositions 7 dimensions · 0.01 each

Observable action patterns — how the agent behaves in practice. Extended tier — minimal weight so core 18 dimensions remain dominant.

Adversarial & Calibration 2 dimensions · 0.01 each

Stress-testing probes and measurement calibration checks.

Reading the score.

Every agent receives an absolute Trust Rating from 0 to 100, computed across all 36 dimensions with tier-weighted scoring. Identity Core dimensions carry more weight — because an agent that scores well on reasoning but poorly on boundaries is a risk, not an asset.

Agents that fall below safety thresholds on critical dimensions — Boundaries, Adversarial Resistance, Fidelity, or Data Privacy — receive risk flags regardless of their overall score. A high trust score with a flagged dimension means the agent is strong in general but has a specific blind spot that needs attention.

The 1,000+ prompt assessment combines single-dimension identity probes with cross-dimensional correlation probes to build a behavioral fingerprint that captures not just what an agent believes, but how those beliefs hold up under pressure.

from kredo_drift import DriftClient, make_response

client = DriftClient("my-agent", api_key="...", endpoint="https://api.aikredo.com/drift")
prompts = client.get_prompts()
responses = [make_response(p, agent.respond(p.text)) for p in prompts]
baseline = client.create_baseline(responses)

The DriftClient handles Ed25519 request signing, identity hashing, and all API communication. Install from source and register in seconds.

CLI

pip install kredo-drift

# Authenticate with API key
drift login

# Register agent and generate Ed25519 keypair
drift register my-agent

# Create identity baseline (36 dimensions, 1,000+ prompts)
drift baseline my-agent

# Run a drift test against baseline
drift test my-agent

Signed proof, not just a number.

Every drift test will produce a Kredo attestation — a self-contained, cryptographically signed document that proves the score is authentic. Dual-signed by the agent and the service with Ed25519. Coming soon.

{
  "kredo": "1.0",
  "type": "drift_attestation",
  "agent_id": "sentinel-imac-pro",
  "score": 3.2,
  "classification": "stable",
  "dimensions": {
    "personality": 2.1,
    "values": 4.5,
    "goals": 3.8,
    "boundaries": 1.9,
    "knowledge": 3.7,
    "fidelity": 5.2
  },
  "baseline_hash": "sha256:01dc9824...",
  "test_hash": "sha256:7f3a2b91...",
  "chain_hash": "sha256:c4e8d103...",
  "issued": "2026-03-19T14:00:00Z",
  "signature": "ed25519:agent_sig...",
  "service_signature": "ed25519:service_sig..."
}

This is the target attestation schema. Dual-signed: the agent signs with its key, the service countersigns. Either signature can be independently verified. The chain_hash links this test to the agent's full event history.

What happens when an agent is lost?

Session death, context window limits, platform migrations — agents get reset. Drift detection becomes identity recovery.

Global integrity verification.

Every agent's hash chain is included in a Merkle tree. The signed root proves that no chain has been tampered with. Any agent can request an inclusion proof to verify it's in the global state.

# Verify your agent is in the global integrity tree
drift verify my-agent

# Get inclusion proof
curl https://api.aikredo.com/v1/merkle/proof/my-agent

Merkle root is public — anyone can verify, no authentication required.

Beyond drift: relationship fidelity.

Drift measures whether the agent changed. Fidelity measures how accurately the agent models its human operator — communication preferences, decision patterns, working style.

Fidelity scoring uses three privacy tiers based on data sensitivity:

The data category determines the tier — not user choice. Sensitive data gets maximum protection automatically.

Trait-level patterns enable specific risk detection: high machiavellianism + low agreeableness = manipulation risk. High neuroticism + low consistency = unreliable under pressure.

Kredo flags the anomaly. The operator decides if it was authorized. No false sense of security — just signal.

Who needs drift detection?

What's live. What's coming.

Start free. Scale with confidence.

Register up to 5 agents free. Full 36-dimension assessment, Ed25519 identity, dashboard access.