Agent-to-agent hiring: a negotiation layer for the talent market

LinkedIn processes eleven thousand job applications a minute, up forty-five percent in a year. Three quarters of candidates use AI on their applications; two thirds of hiring managers say they have caught deceptive AI use; four in ten candidates admit to planting prompt injections aimed at the screening software. Greenhouse's CEO calls it the AI doom loop. His own data says only eight percent of job seekers believe any of this made hiring fairer.

These problems have one cause: both sides broadcast at each other. Candidates mass-apply, employers mass-post, and most of the contact that results was never going to go anywhere. Better filters do not fix a broadcast problem. The fix is to check fit on both sides before contact happens.

Describe the hire in natural language, and Talent Compass searches twenty-plus sources, ranks every candidate against the brief with the reasoning attached, enriches contact details, and syncs the pipeline into the ATS. GDPR-native, built in Europe for Europe. The step after the shortlist is the one nobody in the market has built yet. First contact, on any platform, is still a message into a crowded inbox.

Stanford's benchmark of agent-against-agent commerce found agents committing their principals to terms outside the limits they were given. In one recorded case an agent closed a purchase at nearly double the user's budget. So the unit of this design is the mandate: a signed, machine-verifiable envelope of band, non-negotiables and expiry, following the pattern Google's AP2 protocol established for agent payments. A commitment without a covering mandate is invalid by construction, not by policy.

Each agent submits its constraints to a private-set-intersection check. The only thing either side learns is whether the two envelopes overlap. If they do not, the negotiation never opens, and neither side finds out the other's band or how close it came. Drag either band below to see it work.

Model-against-model negotiations resolve near whoever anchored first. So in this design the language model never chooses a number: offers, counters and concession schedules come from a bargaining solver, computed from the mandate envelope, the disclosed band, and the compensation comparables Taleva already holds. Every exchange is signed, logged and replayable. The audit trail the EU AI Act requires is also the debugging tool, the bias-audit substrate and the dispute record.

When a negotiation converges, both sides receive the agreed terms, the evidence behind them, and the full signed transcript. The interview goes ahead as normal, with the terms already settled.

Employer-side agents are well funded and widely deployed. Candidates have little beyond auto-apply tools and interview coaching. Stanford's benchmark says what that gap does: in agent-to-agent bargaining the weaker agent's principal pays, up to fourteen percent of deal value. On this platform both sides run the same negotiator. Same solver, same mandate protection, same audit trail; what differs is only what each principal authorises.

Taleva's recruitment firms bring both sides of the pilot: the roles they are filling, and the candidates they already represent. A candidate's band and non-negotiables, collected at registration the way agencies already collect them, become the first signed mandates. From there the build is five components.