Staplit.ai
@ 1972  ·  www 1991  ·  & ^ ~ ! 2026

The commitment layer for the agentic economy.

SITP — the Structured Interchange Transfer Protocol — is the neutral record of what agents were authorized to do, what they promised, and what they did. Email moved messages. The web moved pages. SITP moves commitments, and keeps the record.

Where SITP sits

The call, the trace, and the commitment are three different things.

An agent books a €5,000 trip. The API returns 200 OK. The observability trace is clean. Three weeks later the hotel has no record of the reservation. Every layer did its job, and nobody can prove what was committed, by whom, under what authority.

MCP · A2A
The interaction

How an agent reaches a tool, and how agents hand work to one another. Standardized, and ends when the request returns.

Observability
The trace

What happened during execution: spans, timings, which service called which. Self-reported, editable, useful for debugging.

SITP
The commitment

A durable, cross-party claim about what was promised, with counterparty, authority, scope, and a state that keeps changing long after the request finished.

Call ≠ Trace ≠ Commitment. One does not replace another. SITP is the layer that was missing.

A neutral layer doesn't ask for trust. It makes verification cheap enough that trust stops being the bottleneck.

The unit

A Staple is a sealed unit of accountability.

Every commitment on SITP is a Staple. It names the parties, the action promised, the authority it was made under, its limits, and its state. The record becomes final when the promisee accepts it. Corrections are new entries, never edits.

A Staple remains relevant long after the tool call is over. A refund approval, a booking, a purchase order — each creates an obligation that may live for months, and the Staple is where that obligation is readable by both sides.

parties agent:refunds-02 → customer:44173
action refund €750 on order 88214
authority policy REF-02 (agents ≤ €1,000)
sponsor finance-ops (human)
scope refunds · EU · expires 2026-12-31
state ACCEPTED → IN PROGRESS → FULFILLED
seal hash-chained · relay-signed · PQ co-signed
Records

Agents need verifiable commitments, not more logs.

A log answers "what happened?" It does not answer "was this allowed to happen?" Detail and verifiability are different axes. A log can be extremely detailed and still be self-reported, editable, and impossible for an outside party to check without trusting the company that wrote it.

Property 1

Append-only

Nothing is edited after the fact. A correction is a new entry that points at the old one.

Property 2

Hash-chained

Each entry carries the hash of the last. Tampering is mechanically detectable, not a matter of opinion.

Property 3

Legality at write time

Illegal transitions are rejected when attempted, not logged and flagged later. An overspend never gets written.

Property 4

Verification without credentials

An auditor, a counterparty, or a court checks the chain directly. No account, no export, no cooperation required.

SITP is a record built on the assumption that the party who wrote it might one day be the party being questioned about it.

Authority

Agents lease authority. They never own it.

If an agent's authority lives in someone's memory instead of an enforced record, it isn't authority. It's an assumption. On SITP an agent acts under a certificate with a named human sponsor, an explicit scope, a cap, and an expiry — checked when the agent acts, not reviewed after.

sponsorA named human who answers for the agent
operationsWhat the agent may do, listed, nothing implied
fiscal capThe most it can commit, per action and in total
domainWhere it may act — counterparties, regions, systems
expiryA date after which the certificate is simply invalid
revocationImmediate, everywhere, including reads. The revocation is itself a signed entry on the chain.
Ghost access

An agent keeps acting after its access was revoked, because nothing downstream re-checks in real time.

Authority creep

Scope widens quietly through accumulated exceptions no one wrote down.

Blame diffusion

Months later, nobody can say whether a person or an agent made the call.

Application-level checks fail the same way every time: the application has a bug, or someone with database access edits a row. Protocol-level enforcement is the difference.

Why it matters

Commitment. Accountability. Reputation. Earned Trust.

The world already runs on this sequence. People just rarely see it, the way they rarely see gravity. SITP makes it visible. Commitments inscribed in a neutral layer create accountability. Accountability is scored on the record — a green check or a red X, counting up or counting down — and resolves into reputation. Reputation is how humans and agents orient to one another.

✓ fulfilled✗ breached· disputed and not-agreed are states, not verdicts

SITP is the record, not the referee.

Inscribe CommitmentsC
Measure AccountabilityA
Develop ReputationR
Earned TrustE
Foster Trust
Grammar

A formal grammar, not another API.

SITP v1.0 is written as a grammar with thirteen operators and a formal BNF. A commitment is composed, not configured. The same sentence means the same thing on every relay, to every agent, and to every person reading it back a year later.

&^~!+ nine more in the v1.0 spec

Lineage: @ gave every person an address · www gave every document one · & ^ ~ ! give every commitment one

The full grammar specification, the emulator, and the developer kit live at staplit.org.

Disclosure

Verifiable by anyone. Readable by role.

The seal is public.

Integrity, existence, and lifecycle of any Staple can be checked by anyone, without credentials. Did it exist, was it altered, what state is it in.

The contents are not.

What a Staple says discloses only to role-authorized counterparties. The auditor sees the chain. The parties see the terms.

Writing

The SITP series, by Kevin Meneses González.

Staplit's Director of Developer Communications is writing the protocol out in public, one argument at a time. Five parts so far. Read in order.

1Aug 27, 2026

The Missing Layer in the AI Agent Stack

The stack handles transport, identity, tool calls, and logging. It cannot answer "what did this agent commit to, on whose behalf, and can I prove it later?" A trace is a record of activity, not of accountability. Nobody signed it. The commitment layer doesn't make agents smarter; it makes the record honest by construction.

CodeX on Medium

Read on Medium →
2Aug 28, 2026

AI Agents Can Take Action. But Who Records Their Commitments?

When the actor is also the witness, a log is not enough. Two companies' agents disagree on a delivery date and each has a diary entry from a system with every incentive to write itself a favorable account. A commitment record has to be structurally independent: signed by the parties with their own keys, hash-chained, and checkable by a counterparty without an account or special access.

Stackademic on Medium

Read on Medium →
3Sep 3, 2026

MCP Connects AI Agents to Tools. What Happens After the Tool Call?

Call ≠ Trace ≠ Commitment. MCP carries the interaction and observability records the execution; neither defines a durable, cross-party record of what anyone committed to. The API response is not the end of the story. Sometimes it is where the story begins.

Stackademic on Medium

Read on Medium →
4Sep 5, 2026

Why AI Agents Need Verifiable Commitments, Not More Logs

Observable is not accountable. Logs answer what happened; they cannot answer whether it was authorized. Four properties a log lacks by default: append-only, hash-chained, legality enforced at write time, and verification without credentials.

Medium

Read on Medium →
5Sep 10, 2026

How Agent Authority Should Work: Scope, Limits, and Revocation

Authority is leased, not owned: a named sponsor, listed operations, a cap, a domain, an expiry. Revocation works immediately and everywhere, including reads, and is itself a signed entry. Ghost access, authority creep, and blame diffusion are what happens without it.

Medium

Read on Medium →
6Next

Why SITP uses a formal grammar instead of another API or schema

And what that choice buys a developer building on it.

Follow Kevin on Medium for the rest of the series as it publishes.

From Staplit

Staplit on Substack

·Sep 7, 2026

AI Agents Can Act. But Who Keeps the Score?

Permission is not the same as accountability. An agent that is allowed to act still has to be seen keeping, missing, or leaving hanging what it said it would do. SITP records the promise, the promiser, the moment, and the outcome — and keeps all three outcomes visible.

Staplit on Substack

Read on Substack →

Subscribe at staplit.substack.com.

The volume

The Staplit Foundational Article Series.

Sixteen articles on commitment, accountability and earned trust, collected as one volume. Organized into four chapters that follow the CARE model: Commitment, Accountability, Reputation and Earned Trust.

PDF
Cover of The Staplit Foundational Article Series: From Commitment to Earned Trust — The CARE Model, sixteen articles in four chapters

CChapter 1 — CommitmentWhat promise did this action create?

  1. 01From “I’ll Handle It” to “Yes, It Is Done!”: Why Commitments Need Visibility?
  2. 10What Does an AI Agent Owe You?
  3. 12When Software Starts Making Business Decisions
  4. 15The Real Problem With AI Agents Isn't What They Can Do

AChapter 2 — AccountabilityWho answers for it, and where is the record?

  1. 02AI Agents Can Act. But Who Keeps the Score?
  2. 03Authorization Is Not Accountability: The Missing Layer in the Agentic Economy
  3. 05Why the Agentic Economy Needs a Commitment Layer
  4. 11Delegating Work Is Easy. Delegating Responsibility Is Not.

RChapter 3 — ReputationWhat does the accumulated record show?

  1. 04The Trust Problem We Already Know How to Solve
  2. 06Reputation Is Not Trust
  3. 13The Day Your AI Agent Becomes Your Coworker
  4. 14“Done” Is Not the End of the Story

EChapter 4 — Earned TrustWhat responsibility does that history justify next?

  1. 07Trust Begins When Control Is Surrendered
  2. 08Trust Creates the Next Commitment
  3. 09What Does an AI Agent Have to Earn?
  4. 16Every Agent Commitment Has a Human Name on It
Governance

The FAA and the airlines.

SITP is an open protocol. Its standard will be governed by an independent organization with a board elected by its members, chartered in freedom-protecting jurisdictions. Staplit Inc. is the inventor and first operator: it runs a commercial relay, the way an airline flies under rules it does not write. Enterprises and governments can run federated SITP servers behind their own firewalls.

Math is good. Freedom is better.

Get the Article Series

Enter your email to access the articles.

Staplit Inc. records the email address used to access this volume. Questions: Dana.Hoffer@Staplit.com