July 3, 2025

ecosystem.Ai 2026 Roadmap

The 2025–2026 platform roadmap, updated with the innovations that shipped — Runtime 0.9.7, Workbench 2, behavioral algorithms, Two-Tower, and MCP.

Jay

Updated September 2026. We published this roadmap in July 2025. This revision keeps the original intent and records what we actually built on Runtime 0.9.5 → 0.9.7.0 and Workbench 2.0.x.

Introduction

The 2025–2026 plan was not a marketing list. It was a bet that real-time scoring, interaction science, and agent tooling had to live in one prediction platform — not as wrappers on a batch stack. That bet is now in production: a Java Runtime that scores and learns in the closed loop, and a Workbench that configures, pushes, and operates it.

Pins today: Runtime 0.9.7.0, Workbench 2.0.3, Server 0.7.9.0. See the changelog and Runtime release notes.

Next-gen real-time scoring — shipped

The Runtime is still a self-contained scorer. What changed is how much of the platform you can run in one process, and how much of the contract is visible to operators.

• Multi-campaign JVM. Several use cases on a single Runtime instance, with a scheduler and campaign-scoped logs — not one properties file per box.
• OpenAPI + /config. Live configuration dump for support and resilience (0.9.5).
• JIT pre/post-score. Customer Java plugins compile from source in the Runtime; no mandatory plugin build pipeline (0.9.6.0).
• Virtual variables on every deployment type, not only Dynamic Interaction (0.9.6.1, guide).
• External runtime lookups so many scorers can share corpora without exploding Mongo sockets.
• Two-Tower large-catalog retrieval (predictor.model.type=similarity) — embeddings, not a Dynamic Engagement approach. Docs: Two-Tower.
• Runtime Dashboard, campaign Cucumber/BDD, ontology for agents, Kubernetes samples. Hosted demo: medemo.ecosystem.ai .

Closed loop is unchanged and non-negotiable: POST /invocations then POST /response (singular). params is a JSON string.

Interaction science — shipped as algorithms, not slogans

Dynamic Engagement still rolls statistics in the options store and ranks offers on each /invocations call. The innovation is a catalog you can name in config, with model convergence as an operator skill — not a black box.

Dynamic Algorithm approach IDs (do not invent others):

Human Behavioral requires both approach=behaviorAlgos and a sub_approach:

Custom reward plugins, POST /learning, POST /refreshWithOptions, and asymmetric rewards landed on the 0.9.6.1 line. Network Selector (lookup / lookup_passthrough) routes traffic without turning the selector into a bandit.

Full selection tree: Algorithms overview.

Agents, MCP, and Workbench 2 — shipped

The 2025 note promised MCP as a unified tool surface. We shipped three MCP surfaces on purpose — they are not interchangeable (MCP Support):

1. Runtime builtin MCP (POST /mcp on the Java Runtime, 0.9.7+) — score, close the loop, campaign admin from an agent. Dual-era JSON-RPC and 2026-07-28 Streamable HTTP.
2. Workbench MCP — deployments, enrichment, scoring through the operator API.
3. Runtime MCP sidecar — MLFlow refresh and custom Python routers.

Workbench 2.0 is the operator product (platform/versions.json). It pins Runtime; it does not replace it. Campaign manager, runtime push, simulations with visible failures, Spend/Money Personality (including drift), EcoGentic journeys, ontology-backed generation, and Claude/operator agents all sit here.

Generative stack — what landed vs what is still moving

Shipped in the generative configuration:

• Fact injection — prompt library, RAG, and Chat-to-SQL paths that call Runtime APIs.
• Chat to SQL — natural language to SQL against configured stores.
• Vector stores — connectors for retrieval, not a single vendor lock-in.
• MLRun (Train + Deploy) — train/serve path for teams that need a modeling console next to the Runtime.

Still on the path (not claimed as done): Chat-to-SQL V2 extras (schema-diff learning, join/window hints, plan preview), and fully multi-modal generation. We will not pretend those are GA.

Observability and convergence

We did not bolt on a generic APM. We instrumented the learning loop:

• Options store + contacts/responses logging (the data /learning and rolling jobs consume).
• /config and /refresh so Virtual Variables and Dynamic Interaction state can move without a restart.
• Model convergence — priors, levers, and when to switch approach instead of “training longer”.
• Spend / Money Personality drift on the 0.9.7 line.

Grafana-style infra dashboards remain a deployment choice. The platform-native signal is whether scores converge and whether /response is actually closing the loop.

Solutions & modules (current)

How they fit together

Channel  →  POST /invocations  →  Runtime 0.9.7 (plugins + approach)
         ←  ranked offers / uuid
Channel  →  POST /response     →  learning / options store

Workbench 2  configures, pushes, MCP-operates
Server 0.7.9 trains / orchestrates
Ontology     names the same objects across all three

Low-code in Workbench, Java plugins when you need them, Python SDK and notebooks when you are modeling. Prediction as a Service is the product; the versions above are how you pin it.

Why this still matters

• Milliseconds still win the session — see Milliseconds Matter.
• Behavioral intelligence is an architecture, not a plugin — February 2026.
• Named algorithms beat “the AI model”: you can choose lossAversion vs binaryThompson and explain it to risk and marketing.
• Agents that can score (MCP) are only safe if /response stays in the loop.

If you are upgrading: compile pre/post-score plugins against Runtime 0.9.7.0, keep params a string, and never call /responses.