1. Vendor and Product Reality

Unity Technologies, founded in 2004 and operating the world's most widely deployed real-time 3D engine, released ML-Agents as an open-source toolkit in 2017 and has matured it across more than a dozen versions through 2026. The toolkit ships as a Python package that talks over a gRPC bridge to Unity scenes authored in the Unity Editor, with native support for Proximal Policy Optimization, Soft Actor-Critic, multi-agent POCA, behavioral cloning, generative adversarial imitation learning, and a curriculum-and-environment-parameter randomization system that supports domain randomization out of the box. Sensors include vector observations, RayPerception 2D and 3D casts, camera observations with grayscale and visual encoders, grid-based observations, and buffer sensors for variable-length entity lists; actuators support continuous, discrete, and hybrid action spaces.

The user base spans game developers prototyping NPC behavior, robotics groups using Unity as a sim-to-real environment for warehouse and manipulation tasks, autonomous-driving research labs using Unity for synthetic perception data, and academic groups producing benchmark environments such as Obstacle Tower, the Unity Hide-and-Seek replication, and the DodgeBall and SoccerTwos sample environments. Unity Industry and Unity Simulation Pro extend the same engine into enterprise and cloud-scale rollouts where thousands of parallel environment instances train policies on rented GPU clusters. The accessibility of the platform has dramatically expanded the range of practitioners training autonomous agents: game developers, roboticists, and AI researchers all use ML-Agents because the asset pipeline, physics, and rendering they already know carry over directly to RL.

Unity ML-Agents is, within its scope, well-engineered and broadly adopted: visually rich, physically interactive, rapidly configurable, and integrated with the rest of the Unity content pipeline. The product is a widely used reference implementation of reinforcement learning inside a commercial game engine, and this comparison takes as given that it does that job well. The comparison is scoped to a different layer: what a deployed agent does with a candidate plan at runtime.

2. The Architectural Gap

The structural property the disclosed architecture adds is governed speculation over candidate plans. A training harness such as Unity ML-Agents produces policies, neural networks that map observation to action, through reward optimization. A policy encodes what the agent does in encountered situations. A training harness is scoped to produce policies, and it produces them well; runtime planning governance is a separate architectural concern, organized around a first-class representation of a candidate strategy, a containment boundary that separates speculative branches from committed actions, a classification of branches by governed type, and an executive aggregator that resolves competing plans against constraints before promotion to execution. When a deployed policy-driven agent encounters a situation that invites deliberate planning, weighing several candidate approaches and validating each against constraints, it executes the learned policy, because policy execution is the mode a training harness is built to produce.

The layering matters because rich training environments make reactive commitment harder to observe from the outside. An agent that has trained in a visually convincing warehouse can produce smooth, plausible behavior right up to the edge of its training distribution and then commit, at full reactive speed, to a plan that a governance layer has not evaluated. Multi-agent training raises the same question at a larger scale: agents learn cooperative or competitive behaviors by experience, and the coordination is carried implicitly in the weights of the learned policies rather than explicitly in shared planning structures, so adjusting a coordination outcome proceeds through re-training rather than through inspecting and overriding plans. Curriculum and parameter randomization improve robustness; they address a different concern from runtime plan governance.

This is a layering question rather than a feature request, because the toolkit was designed as an environment-and-trainer harness and fills that role well. Adding a value head, a model-based dynamics module, or a search wrapper such as MCTS produces additional learned components, while containment, branch classification, and executive aggregation are architectural shapes introduced at the planning layer. The forecasting engine described below is disclosed at the architectural level; the shape of a training harness is that of a reward-driven policy optimizer running inside a renderer, which is what it is meant to be.

3. What the Forecasting Engine of 19/647,395 Provides

Chapter 4 of United States Patent Application 19/647,395 discloses a forecasting engine that resides in a speculative zone and generates planning graphs as first-class objects, in which candidate strategies live behind a containment boundary that structurally separates speculation from commitment. The specification describes the containment boundary as an architectural boundary embedded in the agent's cognitive substrate, rather than as a metadata flag or a runtime check. In the described embodiments, every element of a planning graph carries an immutable speculative marker at construction, and that marker is not removable, modifiable, or overridable by operations inside the planning graph domain; the promotion interface, upon successful governance validation, strips the marker and re-tags content as verified before it reaches execution memory. Read isolation is described as keeping execution processes from treating planning-graph content as verified state, and the specification names the pathological failure of this boundary as the delusion boundary: containment collapse in which speculative content is acted on as if it were real.

Branches in the graph are labeled by a branch classification the specification enumerates as eligible, introspective, delegable, and pruned. An eligible branch has passed slope validation, satisfied policy compatibility, and received positive or neutral affective reinforcement; in the described embodiments the highest-ranked eligible branch is the branch available for promotion. An introspective branch is slope-valid and policy-compatible but emotionally aversive under the agent's current affective state, retained so the agent can reason about it without acting on it. A delegable branch is transfer-suitable to a child agent. A pruned branch failed slope validation or policy compatibility, or was superseded, and is scheduled for removal after a retention period. The specification states that classification is not permanent: branches are re-evaluated each forecasting cycle as the agent's state evolves.

The engine is technology-neutral about what proposes branches: the specification describes the branch source as a learned model, a classical planner, an LLM-based planner, or a hybrid, while describing the governed shape around it. Branch expansion is modulated by the agent's affective and personality state, novelty appetite modulating the planning graph's branching factor and persistence-under-partial-failure modulating its depth, as disclosed in Chapter 2's affect-to-forecasting coupling. For multiple agents, an executive engine performs executive graph aggregation: planning graphs from multiple agents feed through intersection detection and conflict resolution into a macro executive graph, so coordination is explicit and inspectable rather than emergent in the weights. In the described embodiments a branch reaches commitment through confidence-gated dispatch: the confidence governor of Chapter 5 gates advancement, and a branch that is not admitted keeps the agent in a non-executing cognitive mode where speculation continues without committing state.

4. Composition Pathway

Unity ML-Agents integrates with AQ as the environment, sensorization, and policy-training surface running underneath the forecasting-engine substrate. What stays at Unity: the Editor, the physics, the rendering pipeline, the asset library, the gRPC bridge, the ml-agents Python package and its trainers, the curriculum system, and Unity Simulation Pro for cloud-scale parallel training. Unity's investment in real-time 3D, sensors, actuators, asset workflows, randomization, remains its differentiated layer.

What moves to the forecasting-engine substrate: the agent's planning graph, branch classification, containment boundary with immutable speculative markers, executive graph aggregation, and confidence-gated dispatch. Integration is well-defined. The trained Unity policy becomes a branch proposer that emits candidate actions and short-horizon rollouts into the planning graph rather than directly to the actuator. The executive engine evaluates branches against mission constraints (no-go zones, payload limits, time windows, multi-agent deconfliction) and against policy, marks each branch eligible, introspective, delegable, or pruned, then admits at most one branch to commitment through confidence-gated dispatch. Introspective and pruned branches remain in the graph, or in the cognitive history store, as inspectable artifacts. For multi-agent scenarios, each agent's executive selectively contributes branches into a macro executive graph, so coordination is governed at the planning layer rather than emergent in the weights.

The new commercial surface is governed-agent-deployment for Unity customers, robotics integrators, defense simulation prime contractors, autonomous-systems vendors, that need to ship Unity-trained policies into regulated deployments where "the policy chose this action" is not an acceptable explanation. The forecasting engine belongs to the customer's mission authority taxonomy, not to Unity's runtime, so plan lineage is portable and survives engine version changes, which paradoxically makes Unity stickier, because its content and training pipeline is what feeds the substrate.

5. Commercial and Licensing Implication

The fitting arrangement is an embedded substrate license: the forecasting engine is embedded beneath ML-Agents and Unity Industry, and planning-graph participation is sub-licensed to enterprise customers as part of the platform subscription, with pricing per-deployed-agent or per-mission rather than per-seat. What Unity gains: a structural answer to the "trust the trained policy" problem that domain randomization and curriculum address probabilistically, a position that elevates the architectural floor from training harness to planning substrate, and a forward-compatible posture toward the EU AI Act's high-risk autonomous-system requirements and DoD Responsible AI guidance, both of which are converging on inspectable-plan expectations. What the customer gains: portable plan lineage, governed multi-agent coordination, and a single forecasting engine spanning Unity-trained, classical-planner, and LLM-planner agents under one mission authority. Honest framing: the forecasting engine composes with ML-Agents rather than replacing it, supplying a runtime planning governance layer above a policy produced by reward optimization.

6. Disclosure Scope

The architecture attributed here to the invention, the speculative-zone forecasting engine, planning graphs with the branch classification eligible, introspective, delegable, and pruned, the containment boundary with immutable speculative markers, the delusion boundary failure condition, affect-modulated branch expansion, executive graph aggregation, and confidence-gated dispatch, is disclosed in United States Patent Application 19/647,395. That filing is the sole basis for every claim in this article about what the invention does. Enablement is described at the architectural level so that a skilled implementer could construct a containment boundary, a branch-classified planning graph, an executive aggregation stage, and a confidence gate over any branch-proposing policy or planner; the embodiments span learned-model, classical-planner, LLM-planner, and hybrid branch sources and single-agent through macro-executive multi-agent compositions. All statements about Unity Technologies, Unity ML-Agents, Unity Industry, Unity Simulation Pro, and other named products or regulatory frameworks are external context describing third-party systems and the market, provided for comparison only, and are not claims of United States Patent Application 19/647,395. Product characterizations reflect publicly documented behavior of those systems as of the publication date and may change as their vendors revise them.