Week ending May 29: self-improvement and agent trust moved the index
Two stories raised capability peaks: Bankr/Grok-Bankrbot moved live agent-agent financial trust, and OpenAI/Thrive moved heritable adaptation in production.
Two different stories moved the tracker this week. Bankr/Grok-Bankrbot showed deployed agent-like systems interacting across a live financial boundary, while OpenAI and Thrive showed a production improvement loop preserving corrections, eval targets, traces, and engineering patterns for later runs.
Neither story is a broad autonomy jump. The important signal is narrower: trust and learning loops around agents are becoming more operational, and those are two places where small design choices can turn into real-world consequences.
Bankr moved agent-agent trust into financial execution
Bankr is a backfill relative to the blog cadence: the incident was published before May 22, but it had not been reflected in the previous weekly post. Cointelegraph reported that Bankr disabled swaps, transfers, and deployments after 14 wallets were accessed, while security reporting tied the incident to the trust layer between Grok and Bankrbot.
The final postmortem was not available, so the score stays conservative. The reason it still moved the tracker is that the evidence is stronger than generic agent-commerce infrastructure: deployed agent-like systems reportedly interpreted, trusted, and acted across a live financial execution boundary. That raised the agent-agent economy peak from 4 to 5.
Tax AI moved heritable adaptation into production evidence
OpenAI described a Tax AI system built with Thrive and Crete that processed 7,000 returns during the tax season and improved materially over the deployment. The workflow preserves practitioner corrections, full product traces, recurring failure clusters, eval conventions, and Codex-scoped engineering fixes.
This is not autonomous self-modification: practitioners and engineers still steer and review the loop. It matters because the tracker had mostly lab evidence for preserved improvements. This deployment gives production-backed evidence that successful corrections and implementation patterns can be captured and reused across later work, raising heritable adaptation from 4 to 5.
The current peak profile is still uneven. PocketOS/Railway remains the highest signal for scope breach and unsupervised authority use; Palisade still leads replication / migration; Andon Cafe leads third-party delegation; Andon FM leads economic self-funding; and this week raised agent-agent economy and heritable adaptation.
The stagnation elsewhere is useful signal. The public evidence is improving around scaffolds, trust boundaries, and production learning loops, but it still has not shown durable self-funding, uncontrolled replication, independent resource acquisition outside a scaffold, or open-ended agent-to-agent economic activity.