Keep reviewed team SQL from disappearing into chat threads

An AI SQL workforce is a team of skills whose capabilities are compiled from queries the organisation already trusts. Each agent is a router into a library of reviewed, reusable SQL skills, rather than a raw language model pointed at a database. The workforce grows as the skill library grows; it does not grow because a vendor shipped a new model.

The category exists because the single-agent framing, see what makes one SQL agent production-ready, answers the question "can I trust this one agent" but not "how do twenty teams share what they have already validated." The workforce framing answers the second.

The core wager of SQL agent skills is that the saved and reviewed skill, not the model, is the durable asset. Models change every six months; the SELECT a senior analyst wrote and finance signed off on does not. A workforce composed of compiled skills survives model upgrades, vendor switches, and team turnover. A workforce that is "the model plus prompts" survives none of them.

Most skills, including hand-written Claude Code skills, teach the model how to write SQL. A Chion skill carries the saved and reviewed query itself. The model picks which skill to run; it does not get to rewrite the SELECT underneath.

Walk into any data team and ask how MRR is calculated. You will get three answers. One from the analyst, one from the finance lead, one from a dashboard nobody has opened since the last fiscal year close. They will be subtly different, and reconciling them will take a meeting.

This is not a tooling failure. It is a compilation failure. The org has already paid, many times over, to figure out the right query. The SQL exists. It is correct. It is also untracked, unreviewed, and irretrievable to anyone who was not in the original conversation. Tribal knowledge SQL is the most expensive asset a data org owns and the one it treats with the least care.

The standard response is "let's write better documentation" or "let's consolidate dashboards." Both have been tried, in every org, for fifteen years. The reason they fail is the same: documentation and dashboards are not the runtime. The runtime is whatever the agent, or the analyst, actually executes when the question shows up.

The position this page takes, plainly:

The implication is direct. Investing in raw model access is a bet on a moving target. Investing in compiled SQL skills is a bet on the queries the org has already validated. The first depreciates with every model release. The second compounds with every question the org has already answered.

A SQL skills compiler takes a query a human already trusts and emits a compiled enterprise SQL skill: a single, reviewed artifact with four properties. None of them are exotic. All of them are missing from a raw SQL snippet pasted in Slack.

A workforce that runs against production data inherits every governance requirement a single SQL agent has, multiplied by the number of agents. The non-negotiables: credentials live in an encrypted vault and the model never sees the DSN; the agent connects through a read-only role with RLS enforced server-side; every result is row-capped and timeout-bounded; every query writes one audit row, fire-and-forget by design, by service role only; every compiled skill carries provenance (author, reviewer, last-validated date) the way application code carries commit history.

Read-only SELECT. Credentials in an AES-256-GCM vault. Results capped at 1,000 rows. Each executed query leaves an audit row, written fire-and-forget so a logging failure never blocks your answer.

These are not workforce-specific. They are the floor for any agent pointed at production. The detailed trust framework lives in the four production checks; the implementation specifics for Chion live in the CHION.md specification.

The single-agent question is "is this one agent safe and stable enough to point at production." The workforce question is "how do twenty teams share the skills they have already validated without re-implementing them in each team's silo." Both matter. A workforce that is not built on trustworthy single agents is a faster way to make the same mistakes; trustworthy single agents that never compose into a workforce mean every team rewrites every query.

The portability of the underlying skill format is what makes the workforce composable. That thread is picked up on portable SQL skills.

The shortest path to a workforce is one skill. Pick the question the team answers most often, the one that ends up in Slack every week, and compile the SELECT that already answers it correctly. That is one reusable skill. Do it twice and the library has started. Do it for a quarter and the workforce framing becomes the obvious way to talk about what the team has built.

Chion is the compiler and the runtime. Connect a read-only Postgres, validate the queries your analysts already trust, and save them in Studio so they compile into skills the workforce can call. See how Chion compiles queries into skills, or Chion's AI SQL Analyst framework for the analyst-facing surface; the compiled-skill mechanics live in how Chion compiles SQL skills.

• SQL agent skills are saved and reviewed queries compiled into reusable artifacts an agent runs: capabilities defined by skills, not raw models.
• Tribal knowledge SQL, the trusted query that only lives in one head, is the asset most worth compiling.
• SQL skill compilation produces an enterprise SQL skill: triggers, saved and reviewed SELECT, definitions, provenance.
• A SQL skills framework anchors each repeat question on the saved and reviewed SELECT; the model writes the outer query against reviewed logic instead of the raw schema.
• The internal SQL skills library is the marketplace that matters first; cross-org distribution comes later.
• Models depreciate. Compiled SQL skills compound.