December 2025 - Intent Routing and Parallel Graph Execution

Context

December began the move from a single monolithic conversational agent toward an intent-based, multi-agent architecture — the origin of the orchestration line that runs through 2026. The retained R&D is the routing architecture and the parallel-execution design that keeps it fast, together with the experimental-condition substrate and state-integrity work that support it. Only development hygiene and reporting are excluded, listed concisely at the end. December also closes Q4 with genuine production measurements.

Maps to 2026 R&D projects: Adaptive Multi-Tenant Conversation Orchestration (intent routing, parallel graph); Closed-Loop Agent Evaluation and Optimization (measured Q4 impact).

Intent-Based Routing and Parallel Graph Execution

This project has two faces this month — the routing decomposition itself, and the graph arrangement that keeps the added classification off the critical path. Both belong to the same investigation.

Project and lock

The lock (verrou technologique) is that a single monolithic prompt per client handles every intent identically, which caps response quality and makes per-client prompts grow without bound, yet adding an intent-classification step naively puts a second expensive LLM call in series before every answer — paying the latency cost twice. The uncertainty is how to add intent specialization and routing while keeping the expensive classification off the response critical path, on an execution framework (a Pregel-style superstep graph) whose documentation covers only simple linear or branching topologies, not parallel-analysis-with-convergence.

This month's work

The monolithic router was decomposed into two distinct components with different cost profiles: an LLM-based intent classifier, which is expensive, and a deterministic priority-based action router, which is cheap. The classifier was placed in the first superstep alongside the other analysis nodes (knowledge retrieval, interest-signal detection, visitor profiling) so it runs in parallel rather than in series, and the cheap deterministic router sits at the convergence point where the framework guarantees all upstream results are available. This separation — expensive work parallelized, cheap work on the critical path — is the design insight. The routing itself remained a proof-of-concept: every classified intent still routed to the same answer generator, so the structure is in place but specialized handling per intent is not yet differentiated. The investigation also rested on apparatus built the same month: a cascading per-tenant configuration resolver that applies and compares different behavioural conditions per client and per intent without code changes — the substrate the routing and specialization experiments run on — and deserialization handling that kept existing conversation checkpoints valid across the graph restructuring, so multi-turn experiments are not corrupted by the framework evolution they depend on. Only the generic settings-storage and library-upgrade-maintenance parts of these are ordinary.

Results, proof, and next step

The architecture was implemented and validated structurally: tracing confirmed the first-superstep nodes execute concurrently and that the router receives complete context (intent, interest score, retrieved knowledge) across the superstep boundary. The latency benefit of the parallel arrangement was reasoned from the superstep model but not measured against the prior sequential design in production — the comparison is pending. Because the router routes all intents identically so far, no quality or conversion effect from routing is claimed for December. The classifier/router decomposition is the architectural result; the latency improvement is asserted by design and remains to be measured, and the superstep framework itself is prior art used, not invented. Next step (Q1 2026): differentiate handling per intent and run the A/B test the POC was built for.

Measured Business Impact (Q4 production)

The quarter's production measurements, validating the visitor-intelligence and engagement R&D of October–November (the December routing work is structural and not yet reflected here):

Metric	Oct 2025	Nov 2025	Dec 2025
Initial engagement	1.24%	1.73%	1.86%
Engaged conversations (2+ turns)	29.24%	39.27%	46.32%
Conversion (interaction → form)	~2.89% (baseline)	2.89%	3.15%

Attribution stays conservative: engagement to page-aware questioning plus enrichment, conversation depth to suggested answers and follow-ups, conversion to signal-timed demo proposals. These are correlational production trends across a shifting client mix, not controlled experiments.

Prior-Month Results Review (November 2025 shipped work)

November launched a second engagement experiment alongside the conversation-depth work: a controlled comparison of two answer-presentation styles for the chat surface — a chat-bubble layout with conversational titles against a denser, citation-style layout. The hypothesis was that the more conversational framing would raise engagement. The experiment ran on live traffic from early November and concluded in mid-December, so its result matures here, reported as a quasi-randomized A/B test on presentation style.

The result is a clean null. Across roughly 1,060 exposed visitors per arm over thirty-four days, the two layouts were indistinguishable on every engagement metric measured: the share of conversations reaching a second message was 39.3% in both arms (p ≈ 0.99); the average maximum number of turns was 2.04 against 2.02 (p ≈ 0.78); and the rate at which an engaged visitor clicked a call-to-action was 14.1% against 17.0% — a difference in favour of the conversational layout that did not reach significance (p ≈ 0.09). Presentation style, within the range tested, does not move conversation engagement.

The design is observational in the same ways as the other Q4 reads — a modest per-arm cohort, a shifting client mix, and a single presentation dimension varied at once — but the negative finding is the useful one: it tells the team that engagement gains were not going to come from cosmetic answer framing, which redirected effort toward the intent-routing and specialization line this month's work began. The conversational layout was nonetheless adopted as the default, on the non-significant call-to-action trend and product preference rather than on a measured engagement win; the journal records that distinction rather than dressing the adoption up as a result.

Not retained as R&D

mypy enablement, test-suite housekeeping, and JEI documentation preparation — development hygiene and reporting with no nexus to a named experiment.

Next Steps (Q1 2026)

1. Differentiate per-intent handling and run the routing A/B test (continues the orchestration line into January 2026).
2. Modular/composable prompts to replace the monolithic per-client prompt.
3. Measure the parallel-graph latency benefit against the sequential baseline.