Governed agent negotiation
Agents negotiate. People decide.
A shared protocol and runtime lets your agents propose, counter and accept machine-readable terms against one standard. No commitment binds until a named person clears the approval gate.
Watch the film
The cost a growing agent estate already carries.
Per custom integration, yearly
Lack formal agent accountability
Hit an agent security incident
Plan to add more agents
What the runtime provides.
Deterministic negotiation runtime
An offer–counter–acceptance state machine that tracks status and enforces one common lifecycle across every counterparty, so message formats are never reconstructed per relationship.
Machine-readable term templates
A versioned library covering scope, SLAs, liability, data terms and multi-party obligations. Agents negotiate against a shared vocabulary rather than free-form text.
Authority limits and approval gates
Configurable checkpoints where a designated governor must approve commitments above set value, risk or irreversibility thresholds. Below them, agreed terms proceed.
Inline policy-as-code checks
Compliance and policy-as-code checks run during negotiation and flag terms that breach configured rules before acceptance, moving governance out of after-the-fact review.
How a negotiation runs
From offer to a defensible record.
Agent-to-agent negotiation
Two agents from different ventures exchange offers and counteroffers against shared templates within pre-set authority limits, with state tracked throughout. Terms inside every threshold that pass inline checks are recorded as accepted.
Human governor approval
When a proposed commitment exceeds a threshold or is policy-ambiguous, the runtime halts and routes it to the accountable governor. It cannot bind until they sign off, and the decision is logged.
Audit and dispute evidence
Compliance and risk teams retrieve the complete, ordered record of a negotiation as evidence — supporting internal audit, regulatory enquiry and DIFC dispute resolution without reconstructing events.
Counterparty onboarding
A new counterparty integrates once to the protocol instead of a bespoke connection per relationship, then transacts with any other adopting party on the standard.
What sets it apart
Where authority sits, and what Subchain is not.
Not a ledger
Subchain is a human-governed agreement protocol and runtime — not a blockchain, a token or a consensus network. It governs agent authority, approval-gating and liability, the residual risks current deployments leave unresolved.
Autonomy bounded by human-set limits
AI colleagues conduct exchanges within authority limits, thresholds and routing that a named human sets in advance. They surface exceptions for approval rather than resolving them unilaterally.
A neutral shared standard
Subchain owns the protocol runtime and term-template library as a neutral, DIFC-domiciled contracting standard, so counterparties interoperate as a network rather than through one company's private rail.
The argument
The problem underneath, stated plainly.
Two costs appear at once
When one organisation runs many agents, and those agents must deal with agents at other organisations, two costs appear. The first is integration. Each new counterparty relationship is wired by hand, an expense that recurs for every partner and never fully amortises; vendor estimators put a single custom enterprise integration at roughly $50,000 to $150,000 a year including maintenance, and that burden scales badly as relationships multiply.
The second cost is harder to see and harder to defend. As deployment accelerates, most organisations cannot state with confidence what their agents are permitted to commit to, or produce a clean record of who approved a given commitment. Subchain addresses both from one place: agents transact against a common vocabulary and lifecycle, so a relationship is configured once; and authority to bind stays with an accountable person, recorded step by step.
What we can and cannot yet claim
We are deliberate about the wedge. Subchain is built first for enterprises already running agents in external procurement, in jurisdictions where automated contract formation is recognised — the 2005 UNCITRAL Electronic Communications Convention recognises contracts formed by automated systems — but where agent authority, approval-gating and liability allocation remain unsettled. Those are the risks the design governs.
We are equally direct about what is not yet proven. Direct demand for a standalone governed-negotiation layer is currently inferred from the integration pain around it, not yet confirmed by signed design partners or buyer conversations. Producing real negotiations, real approvals and early signal on willingness to pay is the work of this stage. Authority-to-bind, error and liability allocation, and the DIFC-specific treatment of automated contracting remain open legal questions we research before relying on them, rather than matters we claim to have closed. Subchain is domiciled in the DIFC.
What a serious buyer checks first.
Help prove the layer works.
We are signing the first design partners. Bring the external counterparties you need to transact with, and the commitments you must keep under human sign-off.