A commitment is a record. Anchor it like one.
When an agent negotiates on a principal’s behalf, the offer, the acceptance, the settlement, and the policy under which it acted are all evidentiary artifacts. AAM treats every agent-on-agent commitment as a state change that should leave a tamper-evident, third-party-verifiable record. Bonis Knox provides that record, vendor-neutrally, above any marketplace or payment rail.
Agents are starting to transact, not just respond.
Customer-service agents have given way to research agents, then to coding agents, and now to commerce agents that commit principals to prices, quantities, counter-parties, and delivery terms. Each commitment is a legal event whether or not the principal reads the transcript. The shape of the commitment matters more than the friendliness of the chat wrapper around it.
On 2026-04-24, Anthropic published Project Deal — an internal experiment in which 69 participants, each given a $100 budget, instructed AI agents to negotiate against one another across more than 500 listed items. The agents closed 186 deals end-to-end; humans set preferences at the front of the experiment and returned at the end to physically exchange goods. The published outcome emphasized that stronger models produced measurably better economic outcomes for their principals — and that weaker-model users did not perceive that they had been disadvantaged.
That is the canonical shape of agent commerce: a principal who cannot, in real time, evaluate whether the commitment an agent just made on their behalf was sound. The evidentiary remedy is not to slow the agent down. It is to ensure that every commitment leaves a record durable enough to answer the questions that arrive afterward.
Every commitment becomes a Knox event.
Knox already anchors any event with a canonical byte representation. Transactional mutations qualify trivially: each is a discrete, attributable change to a content- addressable artifact.
An agent commits its principal to a bid, ask, or price-quantity proposal. Knox anchors the SHA-256 of the offer artifact, the chain-of-command stamp identifying the agent and policy under which it acted, and the sequence number in the agent's transactional stream.
A counter-party agent accepts an outstanding offer. Knox anchors both sides of the match — the accepted offer hash and the acceptance artifact — so the binding moment is reconstructible from the anchor chain alone.
The transaction settles — funds, goods, or rights move. Knox anchors the settlement record, the rail it traversed (fiat, on-chain, escrow, custody-transfer), and the timestamp at which the obligations were discharged.
A counter-party, regulator, accountant, or principal opens a dispute over a closed transaction. Knox anchors the dispute artifact and links it to the original offer, acceptance, and settlement anchors — so the full record is retrievable even if the marketplace is gone.
A transaction is unwound — refund, claw-back, charge-back, judicial reversal. Knox anchors the reversal artifact and the policy or order that authorized it, leaving an audit trail that distinguishes a legitimate reversal from a silent rewrite of history.
The trading policy itself — price floors, counter-party allow-lists, daily limits — is changed mid-stream. Knox anchors the policy-document commitment so policy drift is observable across time, not only enforceable at the moment.
The questions are predictable. The records should be too.
Once agent commerce scales beyond an internal pilot, the same five questions arrive in every operational seam. Knox primitives produce records architected to answer each one without re-trusting the marketplace, the agent vendor, or either counter-party.
Did the agent actually agree to this?
The offer and acceptance anchors, taken together with their chain-of-command stamps, identify the exact agent, policy, and inputs under which the binding moment occurred. The Bitcoin anchor pins the moment in time.
Was this transaction within policy when it executed?
The policy commitment in effect at offer time is anchored alongside the offer itself. Subsequent policy changes do not retroactively rewrite what the agent was permitted to do at the moment of action.
What was the taxable event, and when?
The settlement anchor records the discharge of obligations in a content-addressed form that survives the marketplace. The taxable moment is independently datable through the public chain.
Who reversed this, and under what authority?
A reversal anchor must reference the policy or order that authorized it. A reversal without a corresponding authorization anchor is itself a finding.
Is the record we hold today the same record that existed then?
The hash chain detects any silent rewrite. The Bitcoin anchor pins both the existence and the order of every commitment to a moment in time independent of any party to the transaction.
What this gives you.
Vendor-neutral
The audit layer is content-addressed, not coupled to any particular marketplace, payment rail, or escrow provider. Operators do not have to choose between the audit layer and the venue they prefer.
Independent verifiability
Anchors are published to the Bitcoin blockchain via OpenTimestamps. Verification does not require Bonis, the marketplace, or either counter-party to be online, in business, or cooperative.
Reversal-resistant history
A legitimate reversal produces its own anchor with an authorization reference. A silent rewrite produces no anchor — and the gap, taken with the hash chain, is detectable from outside the system.
Post-quantum resilient
Transactional commitments may carry post-quantum signatures via Knox Agent #11 Layer 4 — ML-DSA-44 / 65 / 87 (NIST FIPS 204) and SLH-DSA-128s / 192s / 256s (NIST FIPS 205). The audit chain remains verifiable under threat models that assume future quantum-capable adversaries.
Self-authenticating
Every anchor resolves to a court-ready affidavit architected for FRE 902(13) and 902(14) self-authentication. Admissibility in any given matter remains a determination of the presiding court, but the structural requirements are met by construction.
Above any control plane
Whichever control plane authorized the agent and whichever runtime executed the call, the transaction anchor is produced and verified outside that stack. The audit layer is independent, by design, of the layer being audited.
Evidence layer, not enforcement.
The doctrine that governs every Knox surface governs this one. Bonis Systems does not access third-party marketplaces, does not operate counter-agents on those marketplaces, and does not undertake active disruption of any external system. The audit layer is invitational: operators of agents instrument their own transactional mutations through Knox if they want the record. The takedown, when there is one, belongs to lawful authority. Bonis provides the evidence.
The substrate is already shipping.
The same primitives that anchor every other Knox event anchor a transactional mutation. The endpoints, taxonomy, and verifier are public.