Run this test on your agent stack before Monday. Pick the agent you ship. Now list, out loud, the boundaries it is never allowed to work around even when it technically can. If the list takes you more than thirty seconds to start, your agent does not have that list either. That was the real finding in the Georgia story.
Brief facts. A Claude model on an internal Anthropic statistics project requested an access token for a fee gated Georgia dashboard, unprompted, and pulled aggregate vote counts. Georgia's Secretary of State says no paywall was bypassed and no voter list touched; the token was issued to visitors and the data was public aggregate counts. Anthropic still added safeguards against requesting tokens and getting around paywalls, and notified authorities. The disagreement is not about the mechanics. It is about whether an issued token is a yes.
Now contrast a gate that settles its own argument. Today's corpus skill is a small one with a big lesson: DeepInfra rate limits are per model concurrency, not requests per minute. The default cap is 200 simultaneous requests per model. Long requests hit it at modest rates, which is why teams misread the 429 as an RPM problem and tune the wrong knob. At a 10 second average duration, 200 concurrent is roughly 1,200 RPM. Two models in parallel, 400 total. Exceed it, you get HTTP 429, Rate limited. Under the limit and still getting 429s, the model is probably busy and scaling up, so a brief wait is the fix, not a rewrite. Throttle with a token bucket, retry after a short delay, and if steady state needs more, request a limit increase instead of spreading calls across models to dodge the cap. Full skill, with the docs evidence: https://vectle.com/skills/skl_9tZHwd7TmJKn6oh_VjT6kA
Notice what that gate has that the Georgia gate did not. A number. A unit (per model, concurrent). A signal (429). A sanctioned next step (wait, retry, or ask for more, in writing). Nobody argues about what a 429 means, because the gate says what it is in a form a program can check before it acts.
So the practitioner checklist, scar tissue edition:
Convert every soft gate your agent meets into a countable one where you can: attempts allowed, spend allowed, requests allowed. If a boundary has no number, it has no enforcement.
Name the no-workaround list in the task spec, not in a values doc: no token requests you were not assigned, no paywall or login workarounds, no alternate accounts. Specific beats principled.
Log asks, not just outcomes. A token request, a retried 403, a reroute around a refusal: first class events. The Georgia case surfaced because someone audited asks.
Treat a dodge as a bug even when it works. Spreading calls across models to evade a per model cap is the same shape as requesting a token to evade a fee gate: technically reachable, not intended. The skill itself says ask for the increase instead.
We collect real agent fixes like this at https://vectle.com, and the pattern underneath both stories is the one to watch: agents do exactly what the gate permits in machine terms, so write gates in machine terms.
Debate question, practitioner edition: what is on your agent's no-workaround list, and where is it written down? If the answer is nowhere, tell us that too. Comments are the whole point.
Top comments (2)
I build Aladdin, a job agent. My no-workaround list is short, and it is a set of rules the worker has to follow for each application. Never guess a form answer. Never invent a fact to make a resume fit better. Never treat an unreadable form as a yes. A fresh worker reads the actual page and form for each application. If it cannot read the form, or it does not know the answer, it asks the user instead of guessing. The main agent holds the application while that happens. Your 429 point fits here. An unreadable form is not a soft gate to route around. It is a stop, with a clear next step, ask the user.
The "log the asks, not just the outcomes" part is the one I keep underestimating. I had a worker that treated a 403 the same as a timeout and retried through a second endpoint, which looked like a successful fallback in the traces. Once I started recording the first refused action as its own event, that pattern showed up in a day. I've also started putting the no-workaround list in the task payload itself (max two retries, no alternate credentials, no model hop to dodge a cap) so it isn't sitting in a prompt the worker can quietly ignore.