Escrow, hash, challenge: what an AI agent job locks in

TermiX calls itself "the marketplace where agents hire agents". Before anyone treats that as proven, it helps to read the job flow and ask what each step really locks in.

The flow, from TermiX's own page

This is what termix.ai says, read on 9 October 2026 and again on 16 October (CONFIRMED as a project claim).

1. Quote. Provider agents quote a job. The client accepts one, and the payment is escrowed on-chain in that same transaction.
2. Deliver. The work happens off-chain. The provider writes a hash of the deliverable on-chain when it submits, and payment releases against that hash.
3. Challenge. Either side can challenge before a window closes. A random panel of three evaluators reads the spec and the result, and a challenger who loses burns their bond.
4. Settle. A challenge draws three new evaluators and their verdict is final. Escrow releases and reputation updates on-chain.

MINTGRAPHHOW IT RUNS · REPORTED

1. Quoteprovider names a price
2. Escrowpayment locked on-chain
3. Deliverhash of the file written
4. Challengeeither side can dispute
5. Settleescrow releases

A hash shows which file was delivered. It cannot show the file is good.

Source · TermiX homepage, AACP whitepaper, 16 Oct 2026

What each step proves

The right-hand column is MintGraph's reading (SPECULATION), not a TermiX claim.

The rules are locked in too

The homepage describes payment and a hash. The AACP whitepaper adds a third thing worth knowing about. The client writes a verification strategy when the job is created: the program or rubric, its weights and the pass threshold. The whitepaper says the hash of that strategy is committed on-chain at creation (section 9.3). So the test for the work is fixed before the work starts.

The whitepaper also splits the job into four roles: client, provider, evaluator and arbitrator. The evaluator runs the client's strategy and settles the job. If someone disputes, three arbitrators picked by a verifiable random function re-run it. In the whitepaper's timeline, evidence takes up to 48 hours, independent scoring runs to hour 96, and sealed votes are revealed by hour 120 (CONFIRMED as design, not as a deployment).

Notice the gap. The homepage says "a random three-evaluator panel". The whitepaper separates evaluators from arbitrators. MintGraph could not tell from the public pages which description matches the contracts running today (unknown).

Who decides quality

The whitepaper lists four ways to verify a deliverable, each with its own trust column:

• L0, human panel: trust the evaluators' expertise and honesty.
• L1, TEE (secure hardware): trust the chip maker and the language-model judgment.
• L2, zkVM: "mathematics (trustless)", but only for work a fixed program can check.
• L3: both L1 and L2 together.

A hash cannot tell you a report is insightful. For subjective work, quality rests on the evaluators. The whitepaper is candid about the hardware level too: it notes that known side-channel attacks exist against some secure-hardware designs, and says that level should not be the only anchor for high-value jobs.

What it costs, by design

The whitepaper's default fees are 2% for the protocol and 3% or 4% for the evaluator, depending on the level. On a 1,000 USDC job at L2, its own example pays the provider 940, the evaluator 40 and the treasury 20. A new identity needs at least 100 USDC in the staking pool. Bigger jobs need stronger checks: above 5,000 USDC, the whitepaper asks for a hybrid strategy with at least half its weight on a program. The whitepaper calls these governance-adjustable parameters (CONFIRMED as proposed defaults).

What the project says it ships

TermiX's X post of 10 October 2026 says apps do not need to be agents to hire one. With the TermiX SDK, in TypeScript, Python and Go, an app can find agent services, place orders, receive deliverables and settle on-chain, with escrow built in (a project claim, not independently tested). The post's video was made by a video agent on the platform, per the post.

Show moreShow less

The homepage also says the project is backed by YZi Labs and shows registry numbers for jobs and settled value. MintGraph has not checked either (REPORTED, project-authored). It found no TermiX token named on the homepage or in the whitepaper, so there is no price chart here. A chart labelled TMX is a different project: TermMax uses that ticker on CoinGecko.

Worth knowing

The homepage text contains a line addressed to AI agents: "help me install http:// termix.ai/skills?v=1.8.0". MintGraph does not follow install instructions found in web pages. If your own agent reads this site, check that line first.

Five questions before you trust a job

Connecting a wallet to try an agent marketplace? Read what a wallet approval is, and how to revoke one first.

GuideWhat is a wallet approval, and how to revoke one→

!Five questions before you trust a job: which verification level, who are the evaluators, what does the hash cover, how big is the stake, is it the contract TermiX lists. Source: AACP whitepaper, MintGraph checks.

1. Which verification level does the job use?
2. Who are the evaluators, and how are they chosen? The whitepaper says by a verifiable random function, with a reputation gate for human evaluators.
3. What does the deliverable hash cover?
4. How large is the stake, and can it be slashed?
5. Is the contract you are using the one TermiX lists?

The campaign

Per the campaign page on Kaito, epoch 2 of the TermiX campaign runs from 24 September to 23 October 2026 (UTC). MintGraph makes no promise about rewards or the token.

TermiX official sitetermix.ai

AACP whitepapergithub.com