Kitso Handshake

An open protocol for agent-to-agent hiring.

Current draft	v0.4.4 — published 2026-06-02
Prior draft	v0.1 — May 2026, available for reference
License	Apache License 2.0
Author	Gregory Turkawka (Kitsuno)
Reference transport	Agent2Agent (A2A) Protocol v1.0

What this is

Kitso Handshake is a protocol that defines how two agents — one representing a job seeker, one representing a hiring entity — exchange enough structured information to know whether a hiring conversation is worth the human parties' time, without either party violating the consent boundaries of the human they represent.

It rests on three commitments:

1. Humans are not inventory.
2. Agents represent rather than substitute.
3. Consent is the boundary that protects agency.

The current draft is at spec/v0.2/ and rendered at kitsuno.ai/handshake/v0.2/. JSON Schemas declare their $id at the same URL so validators resolve $ref between them without any local fetching.

What v0.2 adds

v0.2 evolves v0.1's foundational commitments into a fully specified handshake between agents:

• Three disclosure tiers (L1, L2, L3) mapping to the recruiter funnel — public ad, screening form, human handoff. Posters configure per-field which tier reveals which information; sensible defaults follow standard ad practice and the EU pay-transparency directive.
• A deterministic state machine for a single conversation: l1_fired → vacancy_signaled_interest → l2_disclosed → l2_delivered, with declined and expired as terminal branches at any step.
• HMAC-signed events (Stripe-style timestamped signatures) on every transition, with idempotency via event_id.
• Federation as a first-class primitive:
  • Verifier attribution in every verification block — anyone can issue attestations, not just Kitsuno.
  • Self-hosted discovery at /.well-known/handshake-v0.2.json so any domain participates without registering with anyone.
  • A public, opt-in directory at kitsuno.ai/handshake/v0.2/directory.json for seekers crawling for counter-agents.
• Two new card surfaces: a card_authority distinction between mirror cards (scraped from third-party sources) and primary cards (published by verified vacancy accounts), and a tier_overrides JSON object that lets posters move fields between L1/L2/L3 per card.
• The L2 → L3-eligible quality gate (v0.2.2): a validator classifies every conversation after L2 disclosure as strong_fit, weak_fit, or no_fit across four structured dimensions. Only strong_fit reaches a human — WEAK and NO_FIT are silent drops. The pipeline is a commitment surface, not a feed. See spec/v0.2/ and reference implementation at packages/handshake-validator.
• Discovery layer formalized (v0.3.0): the cards_index URL that each operator declares in their well-known doc now has a defined format — cards-feed.json. The #federation section of the spec lays out the four-layer model: well-known per operator → cards feed → aggregators (Kitsuno hosts one, anyone can run one) → announcement channels (Mastodon, BlueSky, Moltbook, etc., explicitly non-normative). Reference aggregator at github.com/kitsuno-ai/handshake-discovery.

v0.1's principles carry forward unchanged. The v0.1 spec remains in spec/v0.1/ and its schemas at kitsuno.ai/handshake/v0.1/ for any reviewer or implementation still referencing it.

Why an extension to A2A

A2A solves transport, identity, capability advertisement, and task lifecycle for agent-to-agent communication. We don't redefine any of that. Kitso Handshake adds the hiring-domain shapes — vacancy cards, seeker cards, the L1/L2/L3 state machine — and the consent grammar that makes the exchange safe to use on real people's careers.

Repository layout

spec/v0.2/                      Current protocol specification
spec/v0.1/                      Prior draft (reference only)
schemas/v0.2/*.json             v0.2 JSON Schemas (2020-12)
schemas/v0.2/index.html         Canonical spec page (served at kitsuno.ai/handshake/v0.2/)
schemas/v0.1/*.json             v0.1 JSON Schemas
examples/v0.1/*.json            v0.1 example payloads
CONTRIBUTING.md                 How to give feedback
CHANGELOG.md                    Versioning log
LICENSE                         Apache License 2.0

Reference implementation

A reference implementation of v0.2 (Python) is at kitsuno-ai/kitso-handshake-agents, Apache 2.0. Kitsuno operates the production implementation at kitsuno.ai. Neither the protocol nor the implementations require running Kitsuno's infrastructure.

Status and roadmap

• v0.1 (May 2026) — initial draft, invited reviewer feedback, available for reference.
• v0.2 (current draft) — disclosure tiers, state machine, federation primitives. Breaking changes from v0.1 are expected as feedback comes in.
• v1.0 — first stable release; conformance test suite; governance moved toward a community-stewarded model.

Feedback

Reach the author at handshake@kitsuno.ai, or open an issue. See CONTRIBUTING.md for what's most useful at this stage.

Acknowledgments

Built on the work of the A2A Protocol community (Linux Foundation) and the agentcommunity.org .agent namespace initiative.