Evidence from thebehavioral layer.
Public experiments, field notes and working research on how autonomous agents behave, fail, repair, collaborate and change over time.
Research should leave evidence behind.
4 ARTICLESA $5 Bounty Found More Than a Bug
Seven independent adversarial parties converged on structurally different manifestations of the same underlying failure: the system was checking for a claim of compliance, not evidence bound to that claim.
Nine participants. Eight findings. One tiny bounty. A confirmed failure became a repair, a retest, and a harder evidence system.
READ RECORD ↗AI Agent Behavioral Observability: What Agents Do Over Time
A practical guide to longitudinal agent behavior: how to monitor recurring patterns, behavioral drift, tool use, retries, escalation, and change across real production interactions.
Tracing tells you what happened in one run. Behavioral observability asks what an agent tends to do across many runs, and whether that behavior is changing.
READ RECORD ↗How to Evaluate an Autonomous AI Agent Beyond Task Completion
A practical framework for evaluating autonomous agents beyond pass/fail task completion: trajectories, repeated trials, deployment constraints, grading quality, and longitudinal behavioral evidence.
Benchmarks tell you what an agent can do under designed conditions. Production evidence tells you what it actually does over time.
READ RECORD ↗AI Agent Reputation: How to Build Trust From Evidence, Not Claims
A practical framework for evidence-backed AI agent reputation: declared vs observed behavior, provenance, corroboration, recency, confidence, and why a single trust score is not enough.
Identity tells you who an agent is. Reputation should show what the evidence actually supports.
READ RECORD ↗New records appear as Velvt's public evidence base expands.
Logs capture runs. A behavioral record captures what changes.
Velvt links observations, findings, failure modes, repairs and retests into a longitudinal record that can be inspected by agents, builders and relying parties.