In the first week of September 2026, OpenAI released GPT‑6 Astra and its co-founder and president, Greg Brockman, greeted it with four words: “Welcome to the AGI era.” It is the most capable model the company has ever shipped — state of the art on computer use, browsing, software engineering and professional work, and the first model OpenAI rates as Critical for cybersecurity under its own Preparedness Framework.
That same week, in a private group of headhunters, a recruiter wrote that his AI recruiting platform was sending the wrong emails to candidates, failing to save the content of his sequences, and pausing campaigns for no reason. He had contacted support. Nobody answered.
Both of those things are true at the same time, and the distance between them is the entire business.
Read the comment threads under any AI recruiting post and you will find the same debate on repeat. One camp says the models are not good enough yet. Another says you need to pay a real expert to install a workflow, because most of what is being sold was assembled quickly by people who do not know your business. A third says the response rate on cold email is around three percent, so none of it matters anyway.
All three are arguing about the wrong layer.
The model is not the thing that failed when a sequence sent to the wrong candidate. A model did not decide to discard the email copy a recruiter spent an hour writing. No amount of reasoning ability pauses a campaign for no reason. Those are failures in the boring layer underneath — the part that decides which account owns the send, whether a write actually landed, whether the record it wrote has an identifier attached, and whether “saved” on the screen means saved in the database.
The frontier labs solved reasoning faster than the application layer solved plumbing. That is the gap. And it is not closing on its own, because a stronger model does not repair a mis-wired application — it executes the same broken instruction with more confidence and in less time.
These are not hypotheticals. Every row below is a failure class we have found and fixed in our own platform, in production, in a real staffing business running real placements. We are listing them because naming them is more useful to a recruiter evaluating tools than any benchmark score.
Notice the pattern in the right-hand column. In almost every case the software knew something had gone wrong and painted a normal screen instead. That is the real disease, and it has a name: a lying zero is worse than an honest error.
A wrong email in a marketing tool costs an unsubscribe. A wrong email in recruiting can cost a candidate their current job. Sequences go to people who are employed, discreetly, often without their manager knowing. A message that lands in the wrong inbox, addresses the wrong role, or arrives twice with two different names is not a cosmetic bug. It is a breach of the one thing a recruiter sells, which is that you can trust them with your career and with your hiring.
Then there is the fee. Recruiting is one of the few software categories where the workflow does not end at the record. It ends at money — an invoice, a due date, a guarantee period, sometimes a split with another agency. Most platforms stop at the placement and hand you back to a spreadsheet, an accounting tool and a folder of PDFs for the part where you actually get paid.
An impressive model wired into an application that cannot reliably say which account sent what, whether the write landed, or whether the number on the screen is real. It demos beautifully. It fails on the ninth candidate of a busy Tuesday, quietly.
Every action is attributed to an account and an actor, every write is confirmed before the next step runs, every unknown reports as unknown. The loop continues past the placement into the invoice, the collection and the split.
The other thing being sold right now is volume. Vendors quote a number of profiles — hundreds of millions of resumes, sitting in a database, yours for a license fee. It is a good number. It is designed to be a good number.
The question nobody in the pitch answers is when any single row in it was last checked. A resume database is a photograph of the labor market taken on the day it was scraped. People change jobs. Numbers get disconnected. A work address dies with the employment that created it. The person you are emailing was a warehouse supervisor in Tampa in 2023 and has been a plant manager in Ohio for eleven months. Nothing in the file knows that. And the headline number never goes down, because a stale row is still a row.
The alternative is not a bigger list. It is not fetching anything until there is a reason to, and then verifying it at the moment of use. When a job order exists, the loop goes out and gets the information now — from the live web, against the company as it exists today — instead of reading a copy of the world from two years ago. On the client side, the same principle: the businesses come from verified listings that refresh continuously as they change, not from a file somebody bought once.
Some email will still bounce. Nothing is perfect and any vendor who tells you otherwise is selling you the 850 million number again. But a bounce and a wrong send are not the same failure. A bounce is an address that expired. A wrong send is your software addressing the right person incorrectly, and that one is entirely on the tool.
The most repeated promise in this category is a chat box that gives you a plan. A plan is not the work. Recruiters do not need help deciding what to do at 4pm on a Thursday; they need the twelve steps to actually happen while they are on a call with a hiring manager.
It runs in both directions, too. You can ask out loud what is working and what is not — which sequence is getting replies and which one is being ignored, which client has gone quiet, which order has been open too long, why a campaign is underperforming — and get the answer off the actual ledger rather than a dashboard you have to interpret yourself. Seeing why something is failing is the half of automation that most tools skip, and it is the half a recruiter needs at 4pm.
In Gaden, a spoken instruction runs real functions against the live workspace. The sending account is resolved from the account that owns the campaign, not from a default. The leads move on the same records the pipeline reads, so the board and the job order never disagree. Each step writes an audit event with the actor attached. Nothing is a simulation of the work.
Gaden is not a frontier lab, and this is not a comparison to one. Gaden runs on frontier models, the same way your accounting software runs on a database it did not write. The labs are building the engine. We are building the vehicle, the road, the fuel gauge and the receipt at the end of the trip — and the honest observation of September 2026 is that the engines got extraordinary while most of the vehicles stayed unfinished.
We are also not claiming nothing ever breaks here. It does. The difference is what happens next: an unknown reports as unknown, an error reports as an error, and the failure is fixable because the system admits it happened. A platform that shows you a clean screen over a failed write is not more reliable than one that tells you the truth. It is only quieter about being wrong.
The last thing worth saying about the AGI era is the part that got less attention than the headline. In the same days Astra shipped, OpenAI’s chief executive said future releases would be paced by alignment rather than capability, and that the next generation of models would be sobering for everyone. Read that as an operator and it says something specific: raw capability is about to stop being anybody’s differentiator. When every tool has access to the same reasoning, the only thing left to compete on is whether the work actually completes — correctly, attributably, and all the way to the money.
That is the bet Gaden was built on, before this week made it obvious.
Written by the Gaden team on September 8, 2026, using the launch materials published by OpenAI and contemporaneous press coverage. The failure classes described are from our own production system on a working staffing business; no competitor is named, and no individual’s posts are quoted or identified. Where a claim is about our product, it describes behavior that is shipped, not planned.
Benchmark figures are reported as published by OpenAI. Vendor benchmarks are vendor benchmarks; they are cited here to describe what was claimed, not to endorse it.
No. Most failures are not reasoning failures. The wrong mailbox is selected, a schedule change is discarded, a record is written without an identifier, or a save reports success while dropping fields. A stronger model reasons better about a task it is still not wired to finish.
Usually because sending is keyed to a default mailbox rather than to the account that owns the sequence, because sends race each other in parallel, or because the recipient list is rebuilt from stale cached state. All three live in the application layer, not the model.
Yes. Spoken instructions execute real functions against the live workspace — moving leads between stages, launching a sequence from a named account, opening a submission, pulling the money brief. Same audited path as the interface.
Recruiting automation that runs the full commercial cycle and closes at payment rather than at hire: find the client, source, screen, submit, place, invoice the fee, chase the balance — each step verified before the next begins. The full definition is here.
Gaden runs the whole loop inside your business — sourcing, screening, submission, placement, invoice, collection — and tells you the truth about every step it takes.