Skip to main content
Back to blog

Devotel Orbit agent versioning — where the six shipped surfaces fit together

A follow-up that stitches the routing, evaluation, spend, and value anchors of Orbit's agent versioning story into one post — plus a worked model-bump on a squad specialist.

Orbit Editorial Team

Quick answer

Devotel Orbit's agent surface now carries six versioning checkpoints that were published as standalone topics: per-API-key usage budgets, scoped token exchange, regression pinning, the Practice Studio roleplay surface, the spend governor, and the Agent ROI report. None of them is a planned feature — all are live, tenant-owned controls. This post names each one and shows a version-bump you can run on a single squad specialist in a few minutes.

The six shipped versioning faces

  1. Per-API-key usage budgets — a monthly ceiling per key, with an in-dashboard progress bar and amber/red alerts at 80% and 100%. The place to bound a leaked or runaway integration before the invoice lands.
  2. Token exchange (RFC 8693) — the OAuth surface agents use to mint scoped credentials, bounded to one named resource, with a signed delegation chain you can audit. See the exchange walkthrough.
  3. Regression pinning — saved test conversations pinned to a prompt version id, with a candidate override, so a prompt is replayed against a frozen version before promotion. See the same guide.
  4. Practice Studio — the roleplay surface under Quality → Practice where supervisors author scenarios and agents rehearse, scored by AI. Training a human agent stays separate from promoting an AI one. See the roleplay playbook.
  5. Spend governor / FinOps controls — LLM monthly/daily caps, per-feature thresholds, and a graceful model-downshift block on /insights/llm-spend. See the cost-governance post.
  6. Agent ROI report — per-agent margin read against the attributed value of resolved outcomes, with role-gated visibility. See the ROI attribution guide.

The squads and the practice-studio treatment (multi-agent orchestration) each stood on their own as single-topic posts; this is the named follow-on that stitches them.

A worked version-bump on one squad

Suppose your inbound voice squad routes with a classifier and three specialists — Sales, Billing, Support — as in the squads post. You want to move the Billing specialist from a premium model to a cheaper tier without touching routing.

  1. Bump the model on one member. Open the squad canvas in the dashboard, select Billing, and change its model preset. The classifier and the other two specialists stay exactly where they were; the routing loop guardrails still run on save.
  2. Pin the regression benchmark. Before promoting, take the regression suite for Billing and pin it to the prompt version id currently in production — say pver-014. That freezes the corpus against the shipped prompt instead of the moving live config, exactly as the regression walkthrough describes.
  3. Read the per-stage latency shift. Rerun the suite with the candidate version overriding the pin. The report shows pass rate plus per-set latency versus the previous run, so you see whether the new model shifted response time on the Billing stage — not just whether it still answers. If latency regressed past threshold, hold the promotion and let the rest of the squad keep serving on the old model.

That is the versioning loop in one move: routing unchanged, corpus pinned, latency read per stage, promotion gated on evidence.

Wrap

Each surface above is already live and tenant-owned. The reason this post exists is that the exchange, pinning, practice, spend, and ROI anchors shipped as separate reads; treating them as one versioning surface is the update. Start with the two most relevant: if you run squads, pin your regression suite; if you run human agents, author a practice scenario. From there, budgets and the governor keep spend bounded while ROI reads whether the agent pays for itself.

Frequently asked questions

Do I have to adopt all six versioning surfaces?

No. They are independent tenant-owned controls. A squad on a smaller team often starts with budget plus regression pinning; a contact center with training needs adds Practice Studio; FinOps teams layer spend governance and ROI on top. None of them is required for the others to work.

Does the regression suite touch real traffic when I pin a version?

No. The replay is run in the same sandbox as the studio test panel, scoring a saved corpus against the pinned version id. No live conversations are replayed, and promotion stays with you — the suite only informs the decision.

Devotel Orbit agent versioning — where the six shipped surfaces fit together — Orbit by Devotel