Stage 1 — Observe
The loop starts with cold, deterministic measurement. Tavryn audits every contract approaching its renewal cliff within a 45-day window and asks two questions:- Seat waste: How many provisioned seats are actually active? The agent compares total allocated licences against the count of seats that logged activity in the rolling 90-day window, producing a verifiable dollar waste figure.
- Usage decline: For cloud contracts, is monthly consumption trending down? The agent computes the annualised dollar waste from the observed decline slope.
renew_as_is and the loop ends — saving you from unnecessary renegotiation friction.
Stage 2 — Compare
Once waste is confirmed, the agent evaluates whether renegotiating with the incumbent vendor is smarter than migrating to a competitor. This is the Switch vs. Renegotiate NPV matrix, and it operates entirely on deterministic arithmetic — no LLM guesswork. For each competitor in the vendor catalogue, Tavryn calculates:
The matrix then compares the best competitor’s 2-year net advantage against the incumbent renegotiation savings. A switch recommendation requires all three thresholds to be met: net advantage > $2,000 over 24 months, payback ≤ 12 months, and competitor reputation score ≥ 4.2.
Stage 3 — Negotiate
With the comparison matrix complete, the agent begins a multi-round negotiation against the vendor’s sales representative or API endpoint. Each round follows a structured concession curve:- The agent opens with an anchor price derived from historical deal memory and competitor quotes — never a naive ask.
- A strict walk-away ceiling is loaded from vendor memory. If the vendor’s offer doesn’t cross below this ceiling after the maximum number of rounds (default: 5), the agent exits negotiation rather than accepting a bad deal.
- Competitor quotes are cited explicitly in each round as leverage: “Competitor X is offering equivalent coverage at $Y.”
- Every round and its outcome is written to the immutable audit ledger in real time.
Stage 4 — Review
Before the negotiated deal moves to any spending policy check, an independent Adversarial Reviewer Agent audits it. The Reviewer operates as a dual-pass auditor: it first runs deterministic rule checks, then optionally applies an LLM pass using a different model provider from the primary negotiation agent (reducing correlated failure modes). The Reviewer checks for:- Rushed acceptance — settling in round 1 with less than 15% concession without testing the vendor floor
- Unaddressed seat waste — idle seats above 25% not cited as leverage when concession is under 20%
- Historical underperformance — current discount trailing historical benchmarks by more than 8 percentage points
- Value destruction — any deal where negotiated savings ≤ 0
AGREE
Deal is sound. Usage data and pricing align with vendor benchmarks. Proceed to the policy engine.
CHALLENGE
Deal leaves leverage unexploited. The negotiator gets one additional counter-round citing the specific concern before proceeding.
REJECT
Critical risk detected — negative savings, severe seat waste ignored, or strong historical contradiction. Escalated to a human supervisor immediately.
Stage 5 — Decide
AnAGREE or CHALLENGE-resolved verdict hands control to the Deterministic Policy Engine. This is pure TypeScript mathematics — the LLM cannot influence, override, or approve any transaction at this stage.
The engine evaluates the deal against your organisation’s configured policy parameters and returns one of three decisions: approved, needs_human, or rejected. If the deal exceeds your human_approval_required_above threshold, the engine generates a 1-tap HMAC-SHA256 approval link delivered out-of-band to your designated supervisor.
Approval links expire after 48 hours and are strictly single-use. The policy engine re-runs a full evaluation at the moment of redemption — so a link generated when your budget had headroom cannot be used to authorise a transaction after your budget has been exhausted in the interim.
Stage 6 — Execute
Once the deal is authorised, Tavryn executes the payment in three atomic steps:1
Lock escrow
USDC is transferred into the
ArcEscrow smart contract on Arc Testnet via Circle Developer-Controlled Wallets. The protocol success fee (up to 20% of realised savings) is locked alongside the principal. An on-chain idempotency key prevents any retry from double-funding.2
Verify delivery
The Fulfillment Verification Engine extracts the vendor’s counter-signed order receipt and performs a 4-field document match: price, seats, term, and date must all match the negotiated terms exactly.
3
Release payment
Only after the 4-field match passes does the Verifier role call
release() on the smart contract, sending USDC to the vendor and the protocol fee to the treasury. If verification fails, funds remain locked until a human supervisor resolves the discrepancy.Stage 7 — Learn
The final stage closes the feedback loop. Tavryn records the negotiation outcome into persistent vendor memory and updates the vendor’s reputation score (bounded between 10 and 100):
When a human supervisor overrides a decision — approving a deal the agent would have rejected, or rejecting one it approved — the override reason code is captured in persistent memory. This structured feedback is injected directly into the negotiation context for the same vendor on future runs, so the agent learns your organisation’s procurement preferences without requiring any reconfiguration.