When agents do the work, proving who did it becomes the product
A professional network built on vouching, a hiring check that flags one in twelve applications, and step-up checks that fire only when risk rises: three answers to the same question.
- The Cosign Awards identify 18 people, 13 companies and 18 products across 49 categories, according to the a16z podcast discussion. Self-reported m:…9820
- Findem found a median flag rate of 8% of job applications screened between August and October 2026, with rates ranging from 1% to 49% across accounts. Self-reported m:…2970
- Step-up verification asks for stronger proof of identity only when the risk of a login or action increases, rather than challenging every user. m:…1549
The network that vouches
Cosign is a professional network built for a world in which an AI agent does some of the networking for you. On the a16z podcast this week, the hosts described it as a shift from static resume-based profiles to a system built on reputation and peer-vouching. What matters, in their telling, is the work a person has done and who can vouch for that work. The Cosign Awards are not a leaderboard but a tool to identify a pool of impactful people, the hosts said, and Cosign is a community product at its root, rather than a traditional leaderboard or prize-giving mechanism. m:…9820
The agent angle is the point. AI agents can be connected to Cosign to find people and opportunities, and the podcast describes agents being connected through Cosign's agents page. The Cosign Awards identify 18 people, 13 companies and 18 products across 49 categories, according to the discussion. The pitch, in short, is that reputation becomes portable once someone else's agent can check it. m:…9820
The hiring check
If Cosign is about proving good work, Findem's data is about catching the record that does not add up. Findem analysed job applications screened by its Authenticity Suite between August and October 2026, and found that at the median account, 8% of applications had a suspicious flag. Rates varied widely across accounts, from 1% to 49%, and the middle half of accounts saw between 5% and 13% of applications flagged. Findem reports the median account rather than a single rate across all applications, so no one large employer sets the figure. m:…2970
A flag, Findem stresses, is not a finding of wrongdoing. It means something in an application does not match other sources, such as a company that cannot be verified or employment dates that conflict with a candidate's online professional profiles. Many flagged details have ordinary explanations. The practical meaning for a hiring team is that verification is a review queue, not a verdict: one in twelve applications at the median account needs a human to look closer, and most of those will turn out to be paperwork, not fraud. m:…2970
The login check
The third piece is the check that happens after hiring, every time someone signs in. Step-up verification is a security control that asks a user for stronger proof of identity only when the risk of a login or action goes up. That is the main difference from standard multi-factor authentication, which asks for the same additional factor every single time. It works, Sift's post argues, because it separates the identity question from the risk question. Account takeover usually begins with valid credentials, which means the first login can look completely normal, so the check has to fire at the moment of the sensitive action, not just at login. m:…1549
The triggers are specific. Session signals include new devices, unusual locations, velocity patterns and non-human behaviour, while action signals include password changes, adding new payees, high-value orders and API key creation. The cost of getting this wrong runs both ways: if you apply challenges to every single customer, that cost adds up and trusted customers will get frustrated and leave, and the most common misapplication is triggering additional checks for routine traffic instead of genuinely risky actions. Sift's post also notes that CISA treats FIDO and WebAuthn and certificate-based authentication as the two forms of phishing-resistant multi-factor authentication it recognises, and cites guidance recommending auditable thresholds that record what triggers a step-up, a hold or a denial. m:…1549
One question, three products
The three posts describe one shift from three sides. As agents act for people, the proof layer splits into separate jobs: who a person is and who vouches for them, whether their record checks out, and whether the actor at the keyboard is the account holder. Each product sells an answer to a different failure: an overlooked contributor, a mismatched application, a login that looks normal but is not. None of the three is independent evidence. Cosign's claims come from the a16z podcast discussion itself, Findem's numbers are Findem's own data, and Sift's post is vendor guidance that leans on others, including CISA. What none of them shows is the cost of being wrong: the contributor the vouching system misses, the candidate a flag delays, the legitimate user locked out of their own account. m:…9820 m:…2970 m:…1549
What is still unresolved
- How Cosign's vouching resists gaming is not addressed in the podcast discussion.
- Findem does not report a false-positive rate for its flags in this post.
- The cost of step-up friction for legitimate users is described qualitatively, not measured.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.