| Internet-Draft | Substrate Observation | July 2026 |
| Morrison | Expires 7 January 2027 | [Page] |
This memo articulates a coordination-protocol anti-pattern observed in
cross-tool agentic systems and describes a substrate-observation
alternative that does not require negotiating a wire format between
heterogeneous concurrent sessions of an identity-bound principal. The
memo is Informational. No protocol element is being proposed for
standardisation; the contribution is the opposite, a delineation of
what should NOT be standardised, and why, with a reference to the
substrate-physics primitives that take its place. Companion memos in
the morrison-* family describe the identity primitives this memo
presumes; specifically, this memo relies on the ~handle namespace
established in [IDPRONOUNS] and the per-principal identity substrate
referenced in [IDACCORD].¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 7 January 2027.¶
Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
When a single identity-bound principal operates several agentic sessions concurrently, whether across different tools, different hosts (a workstation, a laptop, a phone), or different organisational contexts (an individual capacity, a workplace capacity, a contracted capacity), those sessions must deconflict their action without stepping on each other's commits, leases, or external-system state.¶
A natural impulse is to standardise a wire protocol for the sessions to exchange peer-state envelopes: "I am here, working on X, holding lease Y until time T". This memo argues such standardisation is structurally unnecessary, would compound interop burden as new agentic tools enter the ecosystem, and would re-centralise an inherently distributed problem on whatever broker the envelope protocol selected.¶
The alternative is substrate observation: each session observes byproducts of its peers' normal operation (filesystem timestamps, kernel-reported socket peer counts, server-emitted connection counts on shared channels) and forms its own local representation of who-else-is-here. No envelope. No wire format. No broker. Reconciliation occurs post-hoc through substrate-physics commitments (filesystem locks, append-only identity logs, economic settlement, organisational identity append-logs), never through a canonical decision. Identity binding of the principal's surfaces themselves is assumed to follow the conventions of [MCPDNS] and [IDCOMMITS]; this memo concerns only the coordination layer above those primitives.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 RFC2119 [RFC8174] when, and only when, they appear in all capitals, as shown here.¶
The following terms are defined for the purposes of this document:¶
Substrate-emitted byproduct. A filesystem or kernel or network-substrate side-effect of an operation undertaken for some purpose other than coordination, observable to other sessions of the same principal without those sessions having transmitted a coordination message.¶
Decay-to-uncertainty. The property that an observation aged beyond a recency threshold transitions to an explicit "uncertain" state, under which the observing session continues to operate, rather than transitioning to an "absent" state under which the observing session blocks.¶
Mutual hallucination. The property that each session of a principal forms its own local representation of concurrent-peer presence from substrate observations, and that no representation is canonical. Divergent representations are reconciled post-hoc through substrate-physics commitments, not through agreement among the sessions themselves.¶
Substrate-physics cascade. The ordered, non-commutative reconciliation pipeline through which divergent local representations resolve to a single durable history. A reference implementation orders the cascade as (a) filesystem-lock arbitration, (b) per-principal append-only identity-log, (c) external operational settlement (cryptographic non-fast-forward rejection, on-chain transaction receipt), and (d) per-organisation append-only identity-log. No stage in the cascade transmits a coordination marker; each stage is a commitment to the substrate, observed identically by every participating session.¶
This memo identifies envelope coordination, the standardisation of a peer-state-exchange wire format across heterogeneous agentic sessions, as structurally inadequate to the cross-tool identity-bound-principal problem. Three failure modes recur:¶
Every additional agentic tool adopting an envelope-coordination standard must negotiate compatibility with every prior tool's version of the standard. Tool families evolve at different cadences; agreement-by-versioning produces a combinatorial maintenance burden borne by the slowest-moving tool's release cycle. Substrate observation has no compatibility surface to negotiate; tools that emit substrate byproducts as a side effect of their normal operation are mutually visible by construction, regardless of release cycle.¶
Envelope-coordination wire formats imply a destination for the envelopes. A broker, whether discovered via DNS, configured per-session, or shipped by a single vendor, accumulates the peer-state of every session that publishes to it. This collapses what is logically a distributed-observation problem onto a single centralised authority, with the predictable consequences for failure-mode (broker down implies coordination down) and trust (broker operator sees every session's purpose).¶
Envelope payloads typically carry an identifier ("session-id", "principal-id", "agent-id") to permit peers to address each other. Such identifiers become a re-identification surface at the wire layer that the underlying identity infrastructure may have explicitly arranged to bound. Substrate byproducts emit no payload, they are simply present in the substrate, and the inference of peer identity is performed locally by each session from substrate-tier credentials it already possesses (kernel SO_PEERCRED, transport-layer authentication on a shared channel, and equivalent). No wire-layer identifier is exposed.¶
The anti-pattern identified here is the standardisation of a peer-state-exchange wire format for the purpose of deconflicting otherwise-independent action. It is NOT a claim against every message a concurrent session may transmit. A constitutive act, one that brings a new shared commitment into being that did not exist until both parties assented, is categorically distinct from a peer-state envelope. A peer-state envelope DESCRIBES action each session decided independently, transmitted only for the convenience of deconfliction; that is the anti-pattern. A constitutive handshake CONSTITUTES a commitment that does not exist until assent is exchanged, and by its nature it must declare its terms and fail closed. Substrate observation is the correct shape for the former and is neither offered for, nor adequate to, the latter. A deliberate handshake is therefore not in tension with this memo: the two occupy the describe/constitute boundary, not two competing answers to a single question.¶
Sessions observe substrate-emitted byproducts. Three reference observables, listed in order of identity-binding strength:¶
Filesystem modification timestamps on per-session journal files produced by tools that journal to disk. Pseudonymous; compute-location is the observing session's local filesystem.¶
Kernel-reported socket peer-credentials (SO_PEERCRED on Unix-domain sockets, equivalent mechanisms on other systems) for sessions mounting a common per-principal daemon. Identity-bound to the principal owning the daemon; compute-location is kernel-mediated, host-local.¶
Server-emitted concurrent-connection counts on a per-principal event channel maintained by the principal's identity infrastructure. Identity-bound to the principal; compute-location is the server emitting the count, with inference performed locally by the subscribing session.¶
None of these observables is a coordination message. Each exists as a byproduct of the observed session's normal operation: writing its journal, mounting its socket, subscribing to its event channel.¶
Each observable class carries an intrinsic identity-binding strength, and a session's local representation records not merely what it observed but at which tier the observation was drawn. The closed-class provenance grammar of [SUBPROV] provides the vocabulary for that annotation. It distinguishes, for example, a filesystem-timestamp observation from a kernel-peer-credential observation from a server-emitted-count observation, so that downstream reconciliation can weight an observation by the identity-binding strength of its source rather than treating all observations as equally authoritative.¶
When sessions' local representations diverge, typically when two sessions independently take an action that affects shared state (a shared filesystem path, a shared external-system resource, a shared organisational artifact), reconciliation proceeds through the substrate-physics cascade defined in Section 2, ordered: filesystem-lock arbitration, per-principal append-only identity-log, external operational settlement, per-organisation append-only identity-log. Each stage is a substrate commitment. No stage transmits a coordination marker; each stage's outcome is itself observable as another substrate byproduct by every participating session.¶
The cascade is non-commutative: the outcome of an earlier stage constrains the admissibility of a later stage's commitments. This property prevents an attacker from partitioning observations across cascade stages to write conflicting commitments simultaneously.¶
The following example illustrates the operating posture of a session that has formed, from substrate observation alone, a belief that a concurrent peer conflicts with its own work, and that resolves that belief without transmitting a coordination envelope.¶
A principal operates several concurrent sessions. One session prepares to modify a shared artifact and forms the hypothesis that a peer is already modifying the same artifact. Under envelope coordination the session would block, or would transmit a query and await a peer response, before proceeding. Under substrate observation it instead reconciles the hypothesis against three observation classes it already possesses.¶
First, the server-emitted concurrent-connection count on the principal's per-principal event channel (Section 4, third observable) establishes how many peer sessions are present. This observable is identity-bound to the principal and is emitted by the principal's server rather than by any single host, so it accounts for peers on hosts other than the observing session's own. A count derived instead from a host-local observable (Section 4, second observable, kernel peer-credentials on a host-local socket) omits off-host peers and MUST NOT be treated as complete.¶
Second, the per-session filesystem journals (Section 4, first observable) of each present peer disclose, as a byproduct of each peer's normal operation, the artifacts that peer is touching. The observing session binds these into a single local representation (Section 2, mutual hallucination) and observes whether any present peer is touching the shared artifact in question.¶
Third, the filesystem presence of an unattended working surface (an orphaned working tree, or a branch with no live session behind it) is a first-observable byproduct that decays to uncertainty (Section 2). It MUST NOT be conflated with a present peer. A conflict hypothesis grounded in a stale filesystem artifact rather than in a live peer is precisely the divergent local representation the cascade reconciles post-hoc; it is not a live conflict to block on.¶
Having reconciled the hypothesis against present, non-stale peers and found no present peer touching the shared artifact, the session proceeds. It does NOT conclude that conflict is impossible: the mutual-hallucination property (Section 2) denies any local representation the status of a canonical decision. It concludes only that no conflict is observed among present peers, which licenses action, not certainty. Any residual conflict, whether a peer that begins touching the artifact after the observation or an off-host peer the observing session could not see, is caught at commit time by the substrate-physics cascade (Section 5): filesystem-lock arbitration, then the per-principal append-only identity-log, then external operational settlement such as a cryptographic non-fast-forward rejection. This distinction is the whole of the memo's thesis: observation licenses a session to act under uncertainty, and the substrate-physics cascade, not a pre-emptive coordination envelope, is what makes the action safe.¶
A reader may ask whether this memo should propose a standardised set of substrate observables and a standardised reconciliation cascade. It does not. The observables identified above are characteristic of POSIX-derived systems running journal-emitting tools, mounting Unix-domain sockets, and subscribing to HTTP-streaming event channels, substrate that is itself standardised in [POSIX], [RFC8441], and similar. No new substrate standardisation is required for the substrate-observation pattern; it composes directly with existing substrate. Where heterogeneous substrate calls for adapter selection (a Windows tool's journal location differs from a POSIX tool's), the adapter is a tool-private implementation detail, not a wire-format negotiation between sessions.¶
This memo's substrate-observation primitive is structurally distinct from each of the prior-art families surveyed below. The contribution of this memo is the joint articulation of why each family is, by construction, inadequate to the identity-bound-principal cross-tool problem the memo describes; it is not a survey for its own sake.¶
Leader-elected consensus PAXOS [RAFT] requires a designated leader, explicit coordination messages, and a single canonical log. Substrate observation has none of these.¶
Conflict-Free Replicated Data Types [CRDT] require a shared mutable data structure and commutative merge operations. Substrate observation has neither; the cascade described in Section 5 is non-commutative.¶
Gossip and epidemic protocols (Demers et al. 1987, [SWIM] and successors) require explicit anti-entropy or update messages transmitted between nodes on a schedule. Substrate byproducts are not anti-entropy payloads; they are unrelated side-effects.¶
Logical clocks [CLOCKS] (Lamport, vector clocks, Interval Tree Clocks) require piggyback of clock state on application messages. Substrate observation does not piggyback on coordination messages because there are none.¶
Distributed snapshots [SNAPSHOTS] require explicit marker messages injected along communication channels. The reconciliation cascade of Section 5 is triggered by independent operational events, not markers.¶
Cryptographically-chained append-only logs (Certificate Transparency [RFC6962], Git object graphs, blockchain ledgers) are each instantiated by the present memo's cascade as one of its stages, not as the whole. Their novelty in the present context is their composition as the second and fourth stages of a non-commutative cascade triggered by byproduct emission, not their chained-log primitive considered alone.¶
Failure detectors (Chandra-Toueg, [SWIM], Lifeguard) output suspect/dead judgements about peers based on heartbeat latency/absence. Substrate observation outputs uncertainty as a first-class terminal operating state; uncertainty is not a transient state on the way to dead, it is the state the system operates under.¶
Lock-free and wait-free data structures require shared memory between threads. Sessions in the present memo do not share memory; they observe substrate-physics surfaces independently.¶
Web Locks API [WEBLOCKS] and analogous intra-runtime mechanisms operate within a single browser instance and rely on message-passing or lock-arbitration provided by the runtime. They do not generalise to the cross-host, cross-tool problem the present memo addresses.¶
The substrate-observation primitive described in this memo, and the anti-protocol posture for which it is the operational realisation, has biological precedent across five phylogenetically independent subjects. This section cites that precedent not as prior art and not as normative grounding, but as background evidence that coordination as observation of substrate-physics (rather than transmission of a coordination envelope) is a realisable and selectable architectural shape, and not a fanciful engineering construct.¶
Four vertebrate subjects converge: African and Asian elephant matriarchal recognition with infrasonic identity calls and multi-decade individual recognition (McComb, Moss, Durant, Baker and Sayialel 2001 "Matriarchs as Repositories of Social Knowledge in African Elephants" Science 292:491-494); corvid (scrub jay, Eurasian jay, New Caledonian crow, American crow) episodic-like cache memory bound on what-where-when-who, theory-of-mind aware re-caching contingent on observer identity, and trans-generational tool tradition transmission (Clayton and Dickinson 1998 Nature 395:272-274; Emery and Clayton 2004 Science 306:1903-1907; Marzluff et al 2010 Animal Behaviour 79:699-707); cetacean (sperm whale, killer whale, bottlenose dolphin, humpback whale) pod- and matriline-distinct dialect drift, decades-long individual signature-whistle recognition, and culture-as-population-substrate (Whitehead 2003 "Sperm Whales: Social Evolution in the Ocean" University of Chicago Press; Garland et al 2011 Current Biology 21:687-691; Bruck 2013 Proceedings of the Royal Society B 280:20131726); and the human Highly Superior Autobiographical Memory phenotype (LePort et al 2012 Neurobiology of Learning and Memory 98:78-92; Mazzoni et al 2019 NeuroImage 200:163-173). A fifth subject, cephalopods, is phylogenetically independent of the four vertebrate subjects at approximately 600 million years of separation: cuttlefish episodic-like memory (Jozet-Alves, Bertin and Clayton 2013 Current Biology 23:R1033-R1035) replicates the corvid scrub-jay finding in an out-of-vertebrate-family taxon; cephalopod chromatophore camouflage emits identity expression as a continuous output of the substrate-observing nervous system rather than as a stored property, with the emission decaying to incoherent noise rather than to a frozen last-known-pattern when substrate input fails (Hanlon and Messenger 2018 "Cephalopod Behaviour" Cambridge University Press); and cephalopod arm-level motor programs continue executing locally-coherent behaviour against their local substrate under experimental dissociation of central-peripheral coordination (Sumbre, Gutfreund, Fiorito, Flash and Hochner 2001 Science 293:1845-1848). The cephalopod evidence is the strongest available form of biological convergence because it establishes that the architectural pattern is independently selectable across the largest evolutionary gap available in animal cognition, and is therefore a substrate-agnostic architectural primitive rather than a vertebrate-specific contingency.¶
The convergent claim across all five subjects is that coherent identity-recognition cognition is realised in biology as a distributed substrate of heterogeneous nodes with role-specific decay characteristics, not as a single store implemented redundantly, and that coordination of action under partial information is achieved through observation of substrate-physics rather than through transmission of a coordination message between observers. The four properties this memo's specification inherits are: (i) memory is distributed across structurally heterogeneous substrates with role-specific decay characteristics (mirrored by the L1-L4 cascade across heterogeneous substrate-physics layers); (ii) storage at the high-salience tier is append-only with the reading surface as a projection (mirrored by the per-principal IdentityLog append at L2 and the per-organisation collective-scope IdentityLog append at L4); (iii) episodes bind heterogeneous modalities with corroboration per-episode rather than per-stream (mirrored by the substrate-observation primitive which binds heterogeneous substrate-emission classes into a single local representation); and (iv) memory carries metadata about who observed it (mirrored by the kernel-attested or server-attested credential of the emitting session as an intrinsic property of each substrate emission).¶
The split-brain-immune posture of the present memo's anti-protocol shape has its strongest biological grounding in cephalopod distributed-control behaviour: when central-peripheral integration breaks, arm-level programs continue executing coherently against their local substrate. The animal still acts, but as eight quasi-independent agents rather than as one. This is precisely the multi-session split-brain operating posture this memo specifies: divergent local representations under uncertainty, with reconciliation post-hoc through the substrate-physics cascade rather than through pre-emptive central coordination.¶
The biological precedent is cited solely as background. No claim of biological prior art is asserted, and no analogy to biological substrate is normative for any conformance requirement of this specification.¶
This memo requires no IANA actions.¶
Substrate observation surfaces three classes of attack absent from envelope-coordination protocols.¶
A peer emits a substrate byproduct then disappears, leaving an aging observation influencing other sessions' representations beyond its operational lifetime. Mitigation is decay-to-uncertainty with a per-substrate-layer eviction floor: observations below threshold are evicted, not retained at vanishing confidence.¶
A peer emits substrate byproducts to some cascade layers but not others, producing divergent local representations across layers that the cascade cannot fully reconcile. Mitigation is per-observer monotonic layer-coverage commitment: an observer's first emission registers its substrate-set, and later emissions outside that set are quarantined before identity-log write.¶
A peer re-emits aged substrate byproducts to refresh observers' confidence in stale state. Mitigation is observation-id-bound decay, where the decay clock is keyed to the observation identifier rather than to wall-clock receipt time.¶
Substrate observables vary in identity-binding strength. The lowest tier (filesystem timestamps, before any identity binding) is pseudonymous: the observer can infer presence but not identity. Implementations SHOULD operate this tier with refusal to emit in cloud-shell environments (where host identity is shared across users), refusal to emit in continuous-integration environments (where emission would be linkable to public workflow metadata), and refusal to enforce locks at this tier (locks require identity binding; pseudonymous observations do not provide it).¶
This memo grew out of internal architectural design work on coordinating concurrent agentic sessions of a single identity-bound principal across heterogeneous tooling. The realisation that substrate observation suffices, and that envelope coordination is the wrong abstraction at the cross-tool layer, is the central insight behind this specification.¶