Put AI teams on real work

Write the brief, set the tools and approvals, and let a scoped team do the work. Every agent runs in its own workspace under a fixed budget, and you see what it produced before anything goes live.

Get in touch See the workflows →
lucitra · engineering 4 agents · 2 workspaces · 1 awaiting you

Active work

refactor-auth
worktree app-2 · 4 files · +182 −64
running
fix-flaky-tests
retries traced to shared fixture
running
cite-check-glp1-review
31 claims · 31 resolved
done.
patient-export-audit
every field linked to its source record
awaiting approval

refactor-auth · session.ts

const session = await resolve(ctx)
- if (user) return grant(user, ALL)
+ if (!policy.allows(user, scope)) return deny()
+ return grant(user, scope)
Merge to main requires your approval Review →

The loop

Mandate. Assign. Review. Approve

Every mandate follows the same four beats, whether the team is shipping a feature, verifying citations, or tracing a healthcare workflow to its sources.

  1. 01 · mandate

    Every task is a mandate

    You write the mandate: the goal, the tools, the budget, the approvals. Task assignment, never a group chat.

  2. 02 · assign

    A scoped team, isolated

    The team stands up in isolated workspaces of its own. Nothing touches your main branch, your production, or your money without you.

  3. 03 · review

    You see the change itself

    Not a summary of the change, but the change. Comments land where you put them and become context the work has to answer.

  4. 04 · approve

    Anything irreversible waits

    Going live, publishing, placing an order, spending money. Each one stops and waits for a person.

Teams

An org chart, enforced in the database

Managers assign; reports execute; recursion is bounded. The chain of command is a constraint, not a prompt.

  • Roles & reporting lines Every agent reports to someone. Break the chain of command and the work stops, by design.
  • Bounded delegation An agent can hand work to another agent, but only so far. The limit is enforced rather than requested.
  • Heartbeats Agents wake, work, and sleep on a schedule you can read: locally today, 24/7 in the cloud.

Governance

Hard limits, not house rules

Budgets are enforced before a run starts. An over-limit agent can't even begin. Everything it does is written down, forever.

Resource governor

LLM usage, wall-clock, API quotas, data access, and workflow-specific limits are checked before a run can start.

policy: task · warn 80% · hard-stop on

Approval gates

Merges, publishing, data exports, binding commitments, and production deploys stop at the same human approval step.

pending → approved | rejected

Kill switch & audit

A global stop that no run can ignore, and an append-only log of every action, cost, and decision, sealed with integrity hashes.

activity_log: append-only · sha256

Workflows

Same loop. Different evidence

Software development and research run through it. So do healthcare workflows, on the same discipline. Every result traces to a source, and a human takes the last step. The loop never changes; the artifact you review does: a diff, a citation set, or an evidence trace.

lucitra · engineering
refactor-auth
worktree app-2 · 4 files · +182 −64
running
fix-flaky-tests
retries traced to shared fixture
running
preview-deploy pr-214
e2e green · lighthouse 97
done.
Merge to main requires your approval Review →

One loop, many workflows. Each is the same engine with a different template: roles, connectors, limits, and the single human gate.

Bring the first repo

Run it yourself, or have us stand it up inside your company and hand it over. Start with one codebase, one research question, or the workflow you have been putting off.

Get in touch