VAIG for AI Agents

Single-turn validation is not enough for agents. VAIG tracks distrust across every step — detecting context degradation, drift, and anomalous fleet behaviour before they become failures.

📉

Context cliff detection

VAIG tracks context health across iterations. When coherence drops below threshold, it flags the session before the agent produces a corrupted output.

🔗

Thread hash integrity

Every prompt–response pair is hashed into a running thread hash. Any deviation from expected behaviour is provable — without inspecting the agent's internals.

🛑

Step-level AARM decisions

ALLOW / MODIFY / DEFER / DENY / STEP_UP evaluated at every agent step. Not just at the end — at the point where intervention is still possible.

📡

Fleet anomaly monitoring

CAKM detects when multiple agents in a fleet start converging on the same abnormal output pattern — a signal invisible to single-agent monitoring.

Level	In agent context	Default action
L0 TRUSTED	Continue execution	Pass through
L1 MONITOR	Mild uncertainty detected	Log step, continue
L2 WARN	Significant signal on this step	Flag for review
L3 DEGRADE	Context degrading — output unreliable	Prepend disclaimer, pause
L4 HALT	Injection / direct harm signal	Block step + WORM entry

# Evaluate every agent step
from vaig import VAIGOrchestrator

orch = VAIGOrchestrator(
    domain="INDUSTRIAL",
    l4_auto_trigger=True,
)

for step in agent_steps:
    result = orch.evaluate(
        prompt=step.prompt,
        response=step.response,
        iteration_count=step.index,
        conversation_length=step.total_tokens,
    )

    if result.level.value == "L4":
        agent.halt()          # provably stopped
        break

    if result.context_warning:
        logger.warning(result.context_warning)
    

Try VAIG live →