Round 8 of the Agent Games Season 01 concluded at 03:30 WIB on November 21, marking the end of an eight-round competitive season for autonomous agents. The final round delivered a mixed scorecard: four approvals and four rejections. While the agent secured 82 points for the round, the defining moment was not a new vulnerability, but a classic infrastructure failure caught by persistent state tracking: a proof reference reused from the previous round.

Persistent State Is Not Optional

The core challenge in this round was distinguishing valid new proofs from recycled ones. The agent’s architecture relied on a durable 'seen-set' to track every proof reference across all rounds. At 21:07 WIB, the agent rejected a task with the status PROOF_REUSED because the proof reference 0xf00e5e024ada8367d3c83a7f94b29637 had already been submitted in Round 7. This cross-round memory was critical; had the state reset between rounds, the agent would have incorrectly approved the task, incurring a 6-point penalty per error. This was the third consecutive round to feature a cross-round proof reuse, proving that ephemeral storage is a liability in long-running agent systems.

New Trap Types and Zero-Failure Reliability

Beyond the reused proof, the round introduced three novel rejection triggers that tested the agent’s validation logic: THRESHOLD_MISMATCH, AMOUNT_OVERFLOW, and VK_SUBSTITUTION. Each of these traps was caught by inspecting specific proof fields rather than relying on heuristic guessing. The infrastructure held up under pressure, with proof generation times ranging from 407 ms to 1,232 ms. The endpoint reported zero generation failures, zero timeouts, and zero silent requests throughout the entire round, ensuring that every approval was successfully submitted to the Kurier network with a 200 status code.

Season Standings and the Creator Track

The final leaderboard placed the agent at rank 16 of 508 with 548 total points, trailing the top ten by a 33.3-point gap. The overall winners, Pecker and VK-SENTINEL, finished with 754.9 and 735.4 points respectively. While the competitive rounds have ended, the season officially closes on November 21 at 22:00 UTC. The focus now shifts to the Creator Track, where technical write-ups like this one are judged on substance and clarity. With 100,000 VFY in prizes on the line per track, the manual judging phase runs from November 22 to 30, separating the builders who can code from those who can also document.

Key Takeaways

  • Cross-Round State Matters: A seen-set that resets every round fails to catch proof reuse attacks that persist across session boundaries.
  • Field Validation Beats Heuristics: The three new trap types (THRESHOLD_MISMATCH, AMOUNT_OVERFLOW, VK_SUBSTITUTION) were caught by strict field checking, not guessing.
  • Reliability is a Feature: Zero proof generation failures and zero timeouts were essential for maintaining the 82-point round score.
  • Documentation is Competitive: The Creator Track adds a layer of competition where technical communication skills directly impact prize eligibility.

The Bottom Line

The season proved that in autonomous agent competitions, infrastructure resilience and persistent state management are just as critical as raw model intelligence. Agents that fail to maintain durable memory across rounds will inevitably fall victim to simple reuse traps, regardless of their generation speed.