Built, Half-Built, and Next

The useful part of a roadmap is the line between what works and what does not, drawn honestly.

V.E.T.S. development runs in four phases. The first is finished, the second is where the work is now, and the third and fourth are planned rather than built. Below, each phase says which it is. Where something is half-built, this page says that too, because a roadmap that describes intentions as capabilities is worth nothing to the person reading it.

Phase 1: the foundation, complete

COMPLETE

The multi-tenant database, the TeamDoc tree, the permission model, the dynamic forms engine, the patient-history chain and the HTML-in-SQL rendering pipeline. This is the layer everything else stands on, and it has been carrying real work for a long time.

The most checkable evidence for that is the schema change log. stbl_Database_Versions holds 53,117 recorded changes across 11,043 distinct database objects, running continuously from 18 December 2007 to 26 July 2026. Not a burst of activity around a launch, and not a gap in the middle. Nearly two decades of continuous maintenance, visible in one query.

The veterinary logic that sits on it is similarly concrete: 309 astp_VETS_* procedures, of which 71 handle patient history alone.

What the finished foundation actually contains →

Phase 2: the AI layer, current focus

CURRENT

Deployed and running: 14 minions across 78 page contexts, a retrieval index of 2,333 documents on Vertex AI Vector Search as vets_deployed_v3, and a supporting layer of 246 astp_AI_* procedures, 42 atbl_AI_* tables, 49 aviw_AI_* views and 11 afnc_AI_* functions.

Every model string is configuration rather than code, read from stbl_System_CodesGlobal through astp_AI_GetModelConfig: gemini-2.5-flash for interactive work, gemini-2.5-flash-lite for batch, gemini-3.1-pro-preview for premium evaluation.

The scheduled work runs on two separate clocks, which are easy to confuse and worth keeping apart:

Hourly

ajob_AI_GCPVectors carries new and changed content into the vector index, on the hour, all day.

Nightly at 02:00

ajob_AI_Intelligence_Pipeline_Daily runs 29 steps in eight phases in about 22 minutes: summaries, keywords, relationships, confidence recalibration, and the Phase 8 capability refresh.

How the deployed AI layer fits together →

What works today, and exactly how far it goes

This is the section most roadmaps leave out. Agent-to-agent delegation is the newest working capability in the platform, and it is worth being precise about its limits, because the gap between what it does and what the phrase suggests is where most AI roadmaps quietly mislead.

What is real. The platform has recorded … agent-to-agent delegations (… silent-tier, … ask-user). … distinct minions have initiated a handoff. Counts load live when you open this page. Each is recorded as a task, so the trail of who handed what to whom is auditable after the fact rather than inferred from logs.

The routing decision is made in astp_AI_Minion_EvaluateHandoff, which scores a candidate transfer. Below 0.70 it declines and you stay where you are. At 0.90 and above, astp_AI_A2A_HandleHandoff delegates silently; between the two it offers you a choice instead of acting.

What is not real, despite what similar systems claim. astp_AI_A2A_HandleHandoff evaluates once, branches once and returns. There are no multi-step agent chains and no parallel execution. Delegation is single-hop, full stop. The procedure does guard against the obvious failure mode — it refuses a handoff that would loop back into a delegation already open between the same two minions — but chain orchestration is the next thing to build, not a thing that is running.

The routing table tells a similar story. atbl_AI_MinionRoutingPattern holds 262 patterns across all 14 minions, of which 13 are disabled. It also supports negative-keyword vetoes and entity-scoped patterns, and it currently holds zero of each. The safety valves are built and empty. That is worth saying plainly rather than describing the feature as though the rows existed.

Phase 3: the animal domain

PLANNED

Phase 2 has been practice. The techniques were developed against the platform’s own development documentation, because that was the corpus we had and the one where a wrong answer costs nothing. Phase 3 turns them on the animal domain, where a wrong answer costs a great deal more.

None of the following is built. All of it is planned work.

  • A veterinary protocol knowledge base, curated the way the development knowledge base has been
  • Cross-domain knowledge linking, so a veterinarian’s notes, a farrier’s records and a trainer’s observations resolve to the same animal
  • Expert review workflows with AI assistance
  • Automated quality pipelines for content validation
  • Fine-tuned models for veterinary terminology
  • Synthetic data generation for rare conditions
  • Red-teaming for medical safety guardrails

The last three deserve their label. There is no model fine-tuning in the platform today, no synthetic data generation, and no red-teaming step in any pipeline. They are on this list because a system that will eventually answer clinical questions should be adversarially tested before it does, not because any part of that work has started.

Phase 4: opening it up

FUTURE

  • Multi-step agent orchestration — chains and parallel execution, closing the gap described above
  • Provider independence — today every configured model is a Gemini model, without exception. The configuration layer is already the right shape for this; the alternative providers are simply not wired in
  • A third-party developer API
  • A knowledge marketplace for expert contributions
  • Small-model deployment for field use where bandwidth is poor and a phone is the only device present

Phase 4 is where the platform stops being something we build and starts being something a community maintains. What it costs to get there, and how the finished system earns, is set out separately.

What finishing these phases costs →

Inspect what is already built while you read the roadmap:

Multi-Agent Lab Night Shift Dashboard Prompt Tracer

Where to go next

Four phases covers about five years. The argument for why this is worth building at all runs longer than that, and it is made elsewhere.

Where this goes after Phase 4 →