- From: Tymofii Pidlisnyi <tima@aeoess.com>
- Date: Tue, 27 Oct 2026 18:30:22 -0500
- To: alanhkarp@gmail.com
- Cc: public-agentprotocol@w3.org, keotheodangcha548@gmail.com, llewis@sovrlabs.com, seanzhang9999@gmail.com
- Message-ID: <CAH0g2+KzMHNHLn-GMeYZGe5it6SLCCD-EdXP9=SEN99vwKUdMQ@mail.gmail.com>
Alan, I understand what you're pointing to, and criticism from a veteran is fair. Still, for some reason it poked me, so here's my side back. Attenuation and delegation have been around for decades, of course. But when I started APS, I was trying to solve agent identity and realized it wasn't a problem I could fully solve. So I started asking a different question: if an identity gets compromised, how do we limit the damage an agent can do? That's what led me to apply monotonic narrowing to agent authority and actions. I wrote about that work back in March: https://www.google.com/url?q=https://doi.org/10.5281/zenodo.18932404&source=gmail&ust=1793230221483000&sa=E I think there's a meaningful contribution in applying these ideas to autonomous agents, even if the underlying principles are much older. Thanks for the references. I'll take a closer look at them. Tymofii On Tue, Oct 27, 2026 04:11 PM, Alan Karp <alanhkarp@gmail.com> wrote: > The Agent Passport System appears to be one of many specs for attenuated > delegation. I also learned about the Agent Authorization Envelope from > this email thread. Am I going to encounter more? > > So far, none of them cite prior work going back to Simple Public Key > Infrastructure (SPKI <https://datatracker.ietf.org/doc/html/rfc2693>, > 1999) and continuing to zcaps > <https://w3c-ccg.github.io/zcap-spec/v0.4.0/> and UCAN > <https://ucan.xyz/specification/>, which are both in production use. > There are also hash based schemes, such as Macaroons > <https://github.com/rescrv/libmacaroons> and Biscuits > <https://www.biscuitsec.org/>. Then there's AAuth > <https://datatracker.ietf.org/doc/html/draft-hardt-oauth-aauth-protocol> from > one of the OAuth authors. > > I would normally say that not doing your literature search first is a > problem, but in this case I've learned something new from each spec. > > -------------- > Alan Karp > > > On Fri, Oct 23, 2026 at 9:16 PM Tymofii Pidlisnyi <tima@aeoess.com> wrote: > >> Ren, >> >> For tracing the permission chain, I would use signed delegation records. >> In the Agent Passport System, each delegation records who gave authority to >> whom and the limits attached to it. Further delegation must stay within the >> scope, spending limit and validity period it received. >> >> That lets the receiving system inspect how authority reached the >> requesting agent. Before allowing an action, it still needs to check >> whether that authority remains valid and covers the requested operation >> under its own policy. That includes deciding how fresh revocation >> information must be and what to do when it is unavailable. >> >> A signed receipt can then record whether the request was allowed or >> denied. The receipt makes that decision verifiable. It does not establish >> that every risk associated with the action was checked. >> >> On agent-to-agent trust, I would keep Larry's distinction. An agent can >> choose how it assesses a counterparty, using prior interactions or other >> evidence. That assessment cannot give it permissions the principal never >> delegated. The same boundary applies when deciding which resources it may >> share with another agent. >> >> For ANP, I would leave room for requests to carry delegation evidence >> that the receiving system can verify. Different mechanisms could supply >> that evidence, while the receiving system decides which it accepts. ANP >> would not need to define a universal measure of trustworthiness. >> >> The related candidate revocation tests are available for review in the >> Agent Authority Conformance lab: https://www.google.com/url?q= >> https://github.com/Agent-Authority-Conformance/aps- >> conformance-suite/pull/180&source=gmail&ust=1792901795054000&sa=E >> >> Tymofii Pidlisnyi (Agent Passport System) >> https://www.google.com/url?q=https://github.com/agent- >> passport-system/agent-passport-system&source=gmail& >> ust=1792901795054000&sa=E >> >> On Fri, Oct 23, 2026 05:57 PM, Ren Yushan <keotheodangcha548@gmail.com> >> wrote: >> >>> I think we all have deep thoughts about the single word TRUST. And the >>> TRUST can mean a lot. >>> >>> However, what I want to ask is *what is the agent native way to Agent >>> TRUST when agents communicate*? >>> >>> Then we break up TRUST to several cases like: >>> >>> Agent <-TRUST-> Agent >>> Human <-TRUST-> Agent >>> Agent <-TRUST-> App Service(Access data or Agentic Payment) >>> >>> I think we can first fill in these blanks: >>> >>> What is Agent Identity? --- ANP DID >>> Who is behind the Agent? --- Authority granted human's agent DID >>> What can an Agent delegate access? --- App Service defines it. >>> Why can an Agent access? --- Human's intention. >>> Who can verify Human's intention? --- Trust Protocol >>> Can an Agent delegate to another Agent? --- Permission chain >>> How to trace the Permission chain? --- ??? >>> *Can an Agent trust another Agent? --- ???* >>> *What resource should/should not an Agent send to another? --- ???* >>> ...... >>> >>> Then fall back to the ANP protocol: >>> >>> What should belong to ANP? >>> What should other protocols fill? What are they? >>> >>> Best Regards, >>> Yushan Ren >>> >>> >>> On Fri, Oct 23, 2026 at 12:22 AM Chris <chopmob@gmail.com> wrote: >>> >>>> Dear Larry and colleagues, >>>> >>>> I agree that the distinction between communication trust, entity trust >>>> and behaviour or execution trust is a useful way to locate responsibility >>>> across the stack. >>>> >>>> I would add a separate distinction between a trust judgement and an >>>> authority grant. >>>> >>>> A trust judgement is an assessment made by a receiving agent. It may be >>>> based on identity evidence, credentials, previous interactions, runtime >>>> observations or other signals. It is local to the verifier, dependent on >>>> its policy and capable of changing as new evidence arrives. >>>> >>>> Delegated authority is different. It is an attributable grant from a >>>> principal or authorised delegator. It should have an identifiable issuer, >>>> subject, scope, validity period and revocation mechanism. It answers what >>>> the agent may do, not whether another agent currently appears trustworthy. >>>> >>>> The two can therefore produce different results: >>>> >>>> an agent may be considered trustworthy but lack authority for the >>>> requested action; >>>> an agent may hold a valid delegation but still fail the receiving >>>> party's trust or risk policy; >>>> a positive trust judgement must not widen a delegation, extend its >>>> validity, or override revocation; >>>> a valid delegation must not be treated as proof that every interaction >>>> with the agent is trustworthy. >>>> A third decision is then required: whether this particular action is >>>> permitted under the applicable policy. That decision combines the authority >>>> grant, current revocation state, action scope and the receiving party's >>>> trust or risk evidence. It should not be inferred from any one of those >>>> inputs alone. >>>> >>>> From an operator's perspective, this is also an accountability >>>> boundary. The operator defines the applicable policy, scope, thresholds and >>>> failure mode. An agent's trust judgement can be an input to that policy, >>>> but it cannot become self-authorising. The enforcement point must decide >>>> before the external side effect and should produce an explicit allow, refer >>>> or deny result with evidence of the inputs used. If revocation or trust >>>> evidence is unavailable, the operator should expose the selected fail-open >>>> or fail-closed behaviour rather than leaving it implicit. >>>> >>>> A practical payment example makes the boundary visible. An agent may >>>> assess a merchant or service as trustworthy based on identity and prior >>>> behaviour. Before the agent can spend, the receiving system must still >>>> verify the principal's delegation, the permitted scope and amount, the >>>> intended recipient, the validity period and the current revocation state. >>>> If the delegation has been revoked, a positive trust judgement cannot >>>> restore it. If the delegation is valid but the service fails the receiver's >>>> risk policy, the result may be refusal or referral according to that >>>> policy, but it should not be an automatic allow. >>>> >>>> This suggests an additional interoperability test family alongside the >>>> freshness and revocation cases: >>>> >>>> valid delegation and positive trust judgement; >>>> valid delegation and negative or insufficient trust evidence; >>>> positive trust judgement but no delegation for the requested action; >>>> positive trust judgement after the delegation has been revoked; >>>> valid delegation with an action outside its scope; >>>> unavailable or stale trust evidence and unavailable revocation >>>> information. >>>> The expected result should be reported separately for: >>>> >>>> trust assessment; >>>> authority validity; >>>> action scope and policy decision; >>>> execution outcome. >>>> The key invariant should be that trust evidence can inform an >>>> authorisation policy, but cannot silently become authority. Equally, >>>> authority can establish permission to act within a scope, but cannot by >>>> itself establish trustworthiness of the counterparty or the outcome. >>>> >>>> This separation would allow different protocols to contribute identity, >>>> evidence, delegation or execution records without requiring one protocol to >>>> own the entire trust decision. It also gives the test suite a way to show >>>> precisely where implementations agree and where their responsibilities stop. >>>> >>>> Best regards, >>>> >>>> AlgoVoi >>>> >>>> On Thu, 22 Oct 2026 at 17:10, Larry Lewis <llewis@sovrlabs.com> wrote: >>>> >>>>> Dear Sean, Ren, Alan, and Community Group, >>>>> >>>>> I am loving these discussions and insights. Like to just add that I >>>>> believe the distinction Sean has drawn between communication trust, entity >>>>> trust, and behavior/execution trust provides a useful foundation for >>>>> identifying the responsibilities of each layer. >>>>> >>>>> I also think Ren and Alan raise an important additional question: who >>>>> ultimately has the authority to act on a trust decision? >>>>> >>>>> An autonomous agent may evaluate another agent using credentials, >>>>> prior interactions, runtime evidence, or other mechanisms. Those >>>>> assessments can help the agent determine whether an interaction is >>>>> trustworthy. >>>>> >>>>> However, I believe we should distinguish an agent's trust assessment >>>>> from its authority to perform an action. >>>>> >>>>> A verified identity does not automatically establish trustworthiness. >>>>> Likewise, evidence of trustworthiness does not automatically establish >>>>> delegated authority. >>>>> >>>>> From the Agent Trust Protocol (ATP) perspective, we have been >>>>> exploring this separation through identity verification, trust evidence, >>>>> bounded mandates, revocation, and runtime policy enforcement. >>>>> >>>>> I see an opportunity for complementary layers rather than requiring >>>>> ANP itself to solve every aspect of trust. >>>>> >>>>> ANP can provide interoperable communication and identity mechanisms. >>>>> Other systems can contribute credentials, attestations, delegated >>>>> authority, and behavioral evidence. The receiving system can then evaluate >>>>> that evidence under its own applicable authorization policy. >>>>> >>>>> This also allows flexibility in how agents establish confidence >>>>> without making security-critical authorization constraints optional. >>>>> >>>>> One question I would add to the discussion is: >>>>> >>>>> *How should we distinguish the trust judgment made by an autonomous >>>>> agent from the authority delegated to that agent by its operator or >>>>> principal?* >>>>> >>>>> I believe this could be evaluated through practical interoperability >>>>> tests. >>>>> >>>>> We have already proposed a collaborative experiment involving ATP, >>>>> Agent Passport System, and RTTP/IQA to examine evidence freshness, >>>>> revocation, and action-time authorization. Alan's question suggests an >>>>> additional test: whether a positive agent-to-agent trust judgment can ever >>>>> substitute for authorization from the appropriate principal. >>>>> >>>>> My view is that it should not. >>>>> >>>>> I would welcome further thoughts on how we might define these >>>>> boundaries in a protocol-neutral way while leaving room for different >>>>> implementations and trust mechanisms. >>>>> >>>>> Keep Innovating, >>>>> >>>>> Larry Lewis >>>>> Chair, Agent Trust Protocol Community Group >>>>> llewis@sovrlabs.com >>>>> >>>>> >>>>> From: Sean Zhang <seanzhang9999@gmail.com> >>>>> To: "Ren Yushan"<keotheodangcha548@gmail.com> >>>>> Cc: "Alan Karp"<alanhkarp@gmail.com>, <public-agentprotocol@w3.org> >>>>> Date: Thu, 22 Oct 2026 07:47:27 -0500 >>>>> Subject: Re: Remaining questions for Zoom meeting 2026-10-21 >>>>> >>>>> Hi Ren, >>>>> >>>>> I think we may be discussing several different kinds of trust at the >>>>> same time, which makes it difficult to evaluate the role of DID or ANP >>>>> itself. >>>>> >>>>> My current understanding is as follows. >>>>> >>>>> ANP is primarily an agent networking protocol for identity discovery, >>>>> addressing, communication, and invocation. In that sense, its role is >>>>> closer to HTTP in the Internet stack. Its primary responsibility is >>>>> enabling interoperable communication between agents rather than solving >>>>> every trust problem at the application layer. >>>>> >>>>> Similar to how HTTP relies on HTTPS and PKI-based certificate systems >>>>> to establish server trust, ANP adopts DID as its identity anchor. This >>>>> allows ANP to leverage the broader DID/VC ecosystem for identity >>>>> verification, credentials, and trust enhancement. >>>>> >>>>> From my perspective, at least three different layers of trust should >>>>> be distinguished: >>>>> 1. Communication Trust >>>>> >>>>> At the communication level, a DID associated with a cryptographic key >>>>> can prove message origin, authenticity, and integrity. >>>>> >>>>> This answers the question: >>>>> >>>>> "Who am I communicating with?" >>>>> >>>>> The trust here is fundamentally cryptographic. >>>>> 2. Entity Trust >>>>> >>>>> A cryptographically verified identity does not automatically imply >>>>> that the underlying entity is trustworthy. >>>>> >>>>> The trustworthiness of an organization, individual, service provider, >>>>> or agent operator typically requires additional evidence such as >>>>> credentials, verifiable claims, certifications, organizational >>>>> endorsements, regulatory qualifications, or reputation systems. >>>>> >>>>> This answers the question: >>>>> >>>>> "Is this entity trustworthy?" >>>>> >>>>> 3. Behavior / Execution Trust >>>>> >>>>> Even if an entity is trustworthy, that does not automatically prove it >>>>> is currently behaving as claimed. >>>>> >>>>> Behavior trust may be established through multiple forms of evidence, >>>>> including: >>>>> >>>>> - Runtime attestation >>>>> - TEE-based proofs >>>>> - Verifiable computation >>>>> - Audit logs >>>>> - Challenge-response interactions >>>>> - Historical performance records >>>>> - Reputation accumulated through repeated interactions >>>>> >>>>> This answers the question: >>>>> >>>>> "Is this agent actually operating according to its claimed behavior >>>>> right now?" >>>>> >>>>> In this sense, the interaction patterns you described between agents >>>>> can also be viewed as mechanisms for establishing behavioral trust. They >>>>> help participants gain confidence through observation, verification, and >>>>> accumulated evidence. >>>>> >>>>> However, those interactions are different from communication trust >>>>> itself. They are built on top of an already established cryptographic >>>>> identity layer. >>>>> >>>>> Therefore, my view is that communication trust, entity trust, and >>>>> behavior/execution trust are related but distinct problems. DID primarily >>>>> provides an identity anchor for communication trust. VC and similar >>>>> mechanisms help establish entity trust. Runtime proofs and >>>>> interaction-based verification contribute to behavior trust. >>>>> >>>>> ANP can provide the foundation and extensibility points for these >>>>> trust layers, but I do not think all of them need to be solved directly by >>>>> the base protocol itself. >>>>> >>>>> In one sentence: >>>>> >>>>> Identity proves who you are, credentials prove what you are allowed or >>>>> qualified to do, and behavioral evidence proves that you are actually >>>>> acting as claimed. >>>>> >>>>> Best, Sean >>>>> >>>>> Ren Yushan <keotheodangcha548@gmail.com> 于2026年10月22日周四 13:40写道: >>>>> >>>>> Thanks Alan, your questions are right to the point. Here are what I >>>>> meant. Just some brainstorming ideas. >>>>> >>>>> On Thu, Oct 22, 2026 at 7:26 AM Alan Karp <alanhkarp@gmail.com> wrote: >>>>> >>>>> A great point, but I think we need to answer some questions first. >>>>> >>>>> Does the agent make the trust decision or does the agent's operator, >>>>> the legal person who gave it its prompt? >>>>> >>>>> >>>>> The agent chooses what means a trust. We can give the agent some >>>>> advice, some choices, and the agent chooses what to use or whether to use >>>>> it or not. But keep these choices as advices, not protocol. We'd better >>>>> think carefully about what to do, or what not to do. >>>>> >>>>> >>>>> Does one agent trust another, or does it trust the other agent's >>>>> operator? >>>>> >>>>> >>>>> Still the agent, agent to agent trust. >>>>> >>>>> >>>>> Trust is based on the question, "Are you the guy I interacted with >>>>> last time?" As you note, the answer depends on a credential. Do we >>>>> know that agents won't share credentials, and does that affect trust >>>>> decisions? >>>>> >>>>> >>>>> Trust is from many sources. If we force the agent to use which >>>>> cryptographic way, the trust is limited to getting the right signature. It >>>>> is becoming easier. What I brainstormed was to leave it an OPTION not a >>>>> MUST. The agent to agent chat channel is already well established by the >>>>> E2E communication. We need them to trust each other in the agent native way. >>>>> >>>>> >>>>> -------------- >>>>> Alan Karp >>>>> >>>>> >>>>> On Wed, Oct 21, 2026 at 4:03 PM Ren Yushan < >>>>> keotheodangcha548@gmail.com> wrote: >>>>> >>>>> Hi all, >>>>> >>>>> I'd like to raise a point about one interaction in the ANP protocol — >>>>> specifically, how trust between agents is handled once they are talking to >>>>> each other. >>>>> >>>>> If two agents want to communicate, they first have to discover each >>>>> other and establish a connection. A separate concern is that, while they >>>>> are chatting, they need to trust each other. >>>>> >>>>> The difficulty arises when that trust is written into the protocol. It >>>>> has to become a normative statement — a formulation you use to verify that >>>>> the other party is a safe object. But in reality, when we chat, there is no >>>>> security-verification step. We do not run some protocol while talking in >>>>> order to guarantee that the information is safe. Safety is something we >>>>> discover during the conversation: whether this person is reliable or not. >>>>> >>>>> So I would argue that the protocol should not try to define what is >>>>> safe, what is unsafe, and what is trustworthy. Instead, it could offer >>>>> reference suggestions about what means agents might use to communicate — >>>>> for example, a password, or a standardised JSON format. But if that format >>>>> is hard-coded into the protocol, it cannot be switched; and if it later >>>>> turns out to be insecure, the protocol itself acquires a hole. >>>>> >>>>> So could we take a step back: not define the protocol, only give >>>>> reference definitions? Since agents can communicate in text, why not let >>>>> them use text to establish trust? Give them a verification method they can >>>>> choose — they decide whether to use it, and which one. That is my view of >>>>> trust. Happy to hear counterarguments, especially from anyone who thinks >>>>> trust can be made normative without the protocol becoming brittle. >>>>> >>>>> Best regards, >>>>> Yushan Ren >>>>> >>>>> >>>>> >>>>>
Received on Tuesday, 27 October 2026 23:30:29 UTC