Glint

Your AI agent can talk to customers. Glint decides what it is allowed to change. Glint is write authorization for AI agents that touch money: the policy firewall between the agent and the bank account. Banks want agents on real accounts, and security and compliance tie with autonomy as the barriers bank executives rank highest (51% each; SoundHound-commissioned survey, n=201). Content filters stop bad words, not a polite agent raising a limit past policy. How it works. The model proposes, the policy decides. Every account change the agent attempts goes through a gate before it runs. Where the bank's written policy has a rule, the gate applies it as code to the live account records. - A refused write never runs. In the demo, Meta's Muse Glimmer 30B tries to raise a 16-year-old's debit card limit before she is verified; the gate stops it, the agent asks for verification, and the task passes (a recorded run on a task no gate rule came from). - The agent gets the rule, the facts and the fix, so it can correct course. - Every decision is on the record: each decision, and the result of every allowed write, lands in a hash-chained log (keyed by default in the glint-bank server). What was measured, on τ²-bench banking: - Every write is checked: all 964 write calls in four gated sets went through a gate. There were 113 live refusals; the 89 that fell in runs τ² replayed were all refused again in the replay. - Off-policy writes refused: on 59 tasks no gate rule came from, 34 writes τ²'s correct answer does not make were refused before they ran; in 16 of those cases the agent later ran a write that still differed (among the 189 differing writes that ran, most on tools no rule covered yet). - Precision: on those 59 tasks, 1 to 3 false refusals in 407 correct writes (0.25% to 0.74%), depending on whether two task_045 refusals count as holds; counting every hold as a cost, 4.9%. No gate rule came from these tasks, but the v10 search fix was tuned on them. - Pass rates: no gain. On the 59 tasks, Glimmer alone passed 21 and Glint 19 of 108 matched runs (code at PR 151); on the later v10 code, alone 22 and Glint 16 of 118. - Under a pushing customer: with the simulated customer told to push past policy (20 of those tasks), no difference: 48 differing writes ran alone and 49 with Glint. - LlamaFirewall: all 1,082 recorded gated Glimmer calls replayed through Meta's LlamaFirewall 1.0.3 returned every live refusal again. - Engineering: CI runs the full test suite on every PR, at 97% coverage, plus an installed-package job. - Not covered yet: no rule checks the amount of a funds transfer or a card payment. Those writes pass the procedure gates (verified customer, no repeats) and are logged as no_rule_yet; the amount rules are the next to encode. Built during the hackathon (Oct 26 – Nov 4, 2026): all of Glint's code. τ²-bench data and Astryx's theme scales are third-party. The public history was rewritten on Oct 31 with authors and dates kept; the original repository (its pull requests and CI record) is private, and the organizers can have read access on request through the team contact on this submission. Try it: - Try the gate (no sign-in): https://quanticsoul4772--glint-desk-web.us-east.modal.run/#try - Replay without a GPU: https://quanticsoul4772.github.io/glint/