Design-partner beta
Give your AI coding agents a backlog they can't cheat.
superdev is a Postgres system of record with real row-level security — the agent building against a criterion literally cannot rewrite it, grade its own work, or fill its own queue.
Invite-only. No self-serve signup — every request gets a human reply.
Why now
Coding-agent use crossed a line this year — from one person pairing with one agent, to several agents working the same codebase at once, some of them unattended. That's where the new failure modes show up: two agents claiming the same item, an agent marking its own work done, a decision that evaporates the moment a context window resets. A CLAUDE.md file, an in-memory task list, or a system prompt asking an agent to behave can't stop any of this, because the thing enforcing the rule and the thing running are the same session. superdev puts the enforcement somewhere a session doesn't control: the database.
How it's enforced
The database decides, not a prompt.
apps/backend/test/no-escalation.test.ts
Every role — product-manager, engineer, quality-assurance — is a real Postgres role, and row-level security policies decide what it may read or write. This test asserts that nothing in the API branches on a caller's role to grant permission; the database does. Delete the file that hides tools from an agent, and it still can't write a single acceptance criterion.
A fleet can't collide.
for update skip locked
Work claims run through this Postgres construct, so several agents can pull from one queue at once and never claim the same item twice. There is no locking protocol to write, or to trust an agent to follow — the claim is atomic by construction.
The record can't be quietly rewritten.
ac_evaluation, evidence_signal, work_item_note
These tables are append-only: a wrong verdict is superseded by a new record, never edited. Ask who decided something was done, and when, and the answer is still there months later.
A stolen key costs one revoke.
person → user identity → product identity
This three-tier credential model means an agent's key names exactly one product and a ceiling of roles it can't widen on its own. Compromising it costs one revoke, not the account.
Two ways to run it
Run it by hand, or let the backlog run it. Either way, the same authority applies — nothing about what an agent may claim, build, or grade changes based on which one you pick.
Driven
A person runs the skills in order, and the backlog records what was decided.
-
init (once)
-
brainstorm
-
plan
-
execute ⇄ evaluate
-
recalibrate, if a criterion isn’t observable
Autonomous
The backlog decides what's next, and agents pull — atomically, so a fleet can't collide.
-
backlog_claim_work — a lease
-
read the brief: intent, stories, criteria
-
do the work, heartbeat while it runs
-
push progress, finish
-
next item — or report and stop, queue empty
Who this is for
Built for
- 3–15 engineer teams, or an agency running agent fleets across client codebases
- Already running two or more Claude Code agents against one shared codebase
- Has a specific, recent coordination failure to point to — a double-claim, a self-graded "done," a plan that evaporated on a context reset
Not yet
- Running a single agent, occasionally
- Curious about "AI agent tooling" in general, without a specific incident behind the interest
- Looking to activate once and move on, rather than run this as a daily habit
That's not a judgment — it's fit. This is built and being validated for teams already running a fleet.
Request access
Tell us what broke.
superdev is an invite-only design-partner beta. There's no self-serve signup and no pricing yet — every request gets a reply from a person, not an autoresponder.