Execution Synthesis

Mechanism

The capability envelope is a structured object enumerating the operations the agent's substrate can presently execute, the resource costs of each operation along declared dimensions such as compute, memory, latency, and energy, and the preconditions under which each operation is admissible. The envelope is derived from substrate introspection at agent admission and is refreshed on declared events: substrate migration, capability negotiation with a peer, expiry of a leased capability, or detection of substrate degradation. The envelope is canonical state, recorded in the lineage, and any read of it during synthesis is recorded as a dependency of the resulting plan.

The intent is a structured declaration of the goal state, the constraints under which the goal is to be pursued, and the priority of the constraints. Intent is admitted into agent state through the same proposal-and-commit pipeline that governs other inputs, ensuring that the synthesis function operates only on intents that have passed validation. The synthesis function consumes envelope and intent and produces a plan: a directed structure of operations drawn from the envelope, annotated with the resource costs claimed against each envelope dimension and with the preconditions each operation relies on.

Synthesis is bounded by construction. The plan-generation procedure draws operations only from the envelope, so any plan it produces is by construction within the envelope's set of admissible operations. Resource accounting accumulates costs across the plan and compares the totals against the envelope's per-dimension limits; a plan whose accumulated cost exceeds any limit is rejected at the end of synthesis and triggers re-synthesis under tightened constraints. When tightened constraints admit no plan, the synthesis function emits a non-synthesis verdict, recording in the lineage which dimension was binding and which intent constraint could not be satisfied.

The plan carries a witness consisting of the envelope identifier read during synthesis, the intent identifier consumed, the per-operation envelope citation, the per-dimension resource accounting, and a hash binding the plan to the envelope at the time of synthesis. A verifier presented with the plan and the lineage can confirm that the plan's operations are subsets of the envelope, that the resource accounting is consistent with the plan, and that the envelope referenced is the envelope that was canonical at the time of synthesis.

Operating Parameters

Synthesis is parameterized along several declared dimensions. The first is the granularity of the envelope, ranging from coarse capability classes to fine-grained per-operation descriptors. Coarse envelopes admit faster synthesis and shorter witnesses but produce plans whose resource accounting carries larger error bars. Fine envelopes admit tighter resource accounting at higher synthesis cost. The choice is declared in policy and may be context-dependent, with high-stakes intents triggering synthesis against a finer envelope than low-stakes ones.

The second parameter is the cost model used in resource accounting. The cost model declares how operations compose: whether costs are additive across sequential operations, whether parallel operations share or contend for a dimension, whether costs are deterministic or distributional. A distributional cost model produces plans whose witness includes a confidence bound on the accumulated cost, allowing the verifier to accept or reject the plan against a declared risk tolerance.

The third parameter is the synthesis horizon, declaring how far forward in time the plan must extend before the synthesis function considers the plan complete. Short horizons produce plans that must be re-synthesized frequently and that can adapt to envelope changes; long horizons produce plans that commit to longer trajectories and that require correspondingly more conservative resource accounting against envelope drift. The fourth parameter is the non-synthesis policy: whether non-synthesis emits a request to expand the envelope, a request to relax the intent, or a terminal verdict that the agent cannot proceed.

Alternative Embodiments

The mechanism admits embodiments that vary along the synthesis algorithm and the envelope source. In a search-based embodiment, the synthesis function explores plans by combinatorial search over envelope operations, pruning branches whose accumulated cost exceeds envelope limits. In a constraint-solver embodiment, the synthesis function encodes envelope and intent as constraints and invokes a solver whose certificate of solution serves directly as the witness. In a learned-policy embodiment, the synthesis function consults a model trained to propose plans, and the proposed plan is then validated against the envelope through the same accounting procedure used in other embodiments; the model functions as a proposal source whose output is structurally untrusted until validated.

The envelope source can be embodied as substrate-local introspection, as a negotiated envelope obtained from a peer or scheduler, or as a hybrid that combines local capabilities with leased capabilities obtained for the duration of a specific intent. In the leased embodiment, the lease itself is recorded in the lineage, and the synthesis function is permitted to draw operations from the lease only within its validity window; expiry of the lease invalidates plans that depend on leased operations and triggers re-synthesis or non-synthesis.

A further embodiment treats non-synthesis as a structured negotiation event. When the synthesis function emits non-synthesis, it records which dimension was binding; this record can be consumed by an upstream coordinator that may relax the intent, expand the envelope through capability negotiation, or route the intent to a different agent whose envelope admits a plan. The non-synthesis verdict thereby becomes a signal in a multi-agent coordination protocol rather than a local failure.

Composition with Adjacent Mechanisms

Execution synthesis composes with the proposal-and-commit pipeline that admits intents, with the integrity field that governs which intents are admissible under the agent's normative state, with capability awareness that maintains the envelope, and with the lineage substrate that records both the synthesis event and the resulting plan. Downstream, execution synthesis composes with the executor that consumes the plan and with the confidence governance that weights the agent's reliance on the plan as the substrate evolves.

Composition with these mechanisms ensures that synthesis is not an isolated planning function but a structural transition of the agent. The plan, once synthesized, is part of the agent's canonical state and is subject to the same retention, forgetting, and verification semantics as any other admitted state. The lineage's record of synthesis allows any subsequent observer to reconstruct why the agent acted as it did under a particular envelope and intent.

Distinction from Prior Art

Prior approaches to plan generation in autonomous systems fall into capability-blind and capability-aware categories. Capability-blind planning assumes the substrate can execute any operation in the planner's vocabulary and relies on runtime failure to detect infeasibility; the planner produces plans that may be impossible, and the system tolerates the resulting failures. Capability-aware planning consults a capability description but typically does so as an unverified read, producing plans that claim feasibility without carrying a witness that the claim is structurally accurate.

The mechanism described here is structurally distinct in that it produces a plan whose feasibility against the envelope is verifiable by any party with access to the envelope and the lineage, and in that it treats non-synthesis as a legitimate structural verdict with declared downstream semantics rather than as an exception. The agent does not attempt operations it cannot execute, and the agent's claim of feasibility is auditable independently of execution.

Disclosure Scope

The disclosure covers the synthesis of plans from a capability envelope and a declared intent, the bounding of synthesis to the envelope by construction, the production of a witness that binds the plan to the envelope at the time of synthesis, the structural treatment of non-synthesis as a legitimate verdict, the parameterization by granularity, cost model, and horizon, the alternative embodiments described above, and the composition with capability awareness, integrity governance, and lineage. It covers any system in which an agent's plan carries verifiable evidence that it was synthesized within a declared capability envelope under a declared intent.

Implementations that produce plans without verifiable binding to a capability envelope, or that treat infeasibility as a runtime exception rather than as a structural verdict, fall outside the disclosure. Implementations that draw operations from a vocabulary broader than the envelope and rely on downstream filtering to reject infeasible plans fall outside the disclosure with respect to the structural bound. The structural property is that synthesis is bounded by construction, witnessed by lineage, and verifiable to third parties, and the disclosure scope tracks that property.

The disclosure further covers the verification protocol by which a third party confirms a plan against an envelope. The verifier reads the plan's witness, retrieves the envelope referenced by the witness from the lineage, walks each operation in the plan and confirms that operation lies within the envelope's set of admissible operations, accumulates resource costs against each declared dimension, and confirms that no dimension was exceeded. The verifier's procedure is itself declared in policy so that an agent and its verifier agree, in advance, on what constitutes a valid synthesis. Implementations that record plans without sufficient detail for this verification, or that bind plans to envelopes that are not themselves canonical state, fall outside the disclosure with respect to the verification property.

The disclosure also covers the structural treatment of envelope drift during plan execution. When the envelope changes after synthesis, the executor consults the witness to determine whether the change invalidates the plan; if it does, the executor halts execution and triggers re-synthesis under the updated envelope. The re-synthesis event is recorded in the lineage as a continuation of the original intent, allowing observers to trace how the agent's plan adapted to evolving substrate conditions. Plans that proceed to execution without consulting envelope drift, such that the plan's feasibility claim becomes stale, fall outside the disclosure with respect to the structural bound on synthesis.