Jules

You will like Jules for the backlog: hand it a GitHub issue and it clones the repo into a Google VM, runs your tests, and returns a pull request while you do something else, 15 free tasks a day. The daily trait is asynchrony, it is a colleague you assign to, not a tool you sit with.

Pick it on a personal Google account for the chores you never reach: dependency bumps, small test gaps, the lint you keep postponing. Do not pick it for a team, since paid plans exclude business accounts and nobody wants company code flowing through personal Gmail; pick Kiro there.

The worst thing is opacity. Dependencies come from an optional setup script inside a Google-hosted VM; if the script fails, the agent works blind and the PR arrives with the confidence of one that ran tests. You cannot inspect the VM, so the only evidence is the log the agent shows you.

The consequence: read the setup output before the diff, every time, and treat a green PR from a red environment as untested. A visible environment status on the PR would change this verdict. The meter is the one thing it gets right: tasks per day, not tokens, so a loop costs a task slot, not a bill.

No benchmark is claimed, so nothing is overstated. The documented architecture: 1. Clone into a virtual machine. 2. Generate a plan and hold for approval before any change. 3. Run tests in place and return a pull request. The first two steps are the principled parts, because the plan is the one artifact a human reviews before tokens are spent on code.

The consequence is that quality depends on whether the model honors the approved plan, which is asserted and unmeasured. A quiet observation: the free plan lists Gemini 2.5 Pro while the changelog lists 3.1, so which model is measured depends on which page you read.

The paid tiers are consumer bundles, Google AI Pro at $19.99 and Ultra from $99.99, so Jules has no pricing power of its own; its price is whatever the bundle costs, set by someone selling storage and video. Paid plans excluding business accounts means zero enterprise revenue today, which for a coding agent is the only revenue that compounds.

The likely pivot is a merge into another Google developer product, with the tasks-per-day meter surviving as a quota somewhere else. No acquirer; the parent is the market. Position: free tier for chores, no team process on it, and expect the name to change.

The demo is a pull request that appears during a meeting. Procurement cannot proceed: the usage-limits page states paid plans are limited to individual Google accounts, with Workspace support in development. That means sixty engineers on personal Gmail addresses granting a Google VM access to the company's GitHub organization, which is a policy violation before it is a purchase.

Nothing about SSO, audit or retention matters until the account model does. CI auto-fixing is attractive and unusable for the same reason. Revisit when Workspace accounts are supported and the first question becomes the seat price. Not yet.

Nothing here runs on my hardware. The VM is Google's, the model is Gemini and no other, and MCP means a curated set of servers Google chose, so the extension surface is a menu, not a config file. What I can bend: npm install -g @google/jules gives me a CLI, there is a REST API, and a setup script plus AGENTS.md shape the environment and the instructions.

That is enough for a cron job that files PRs overnight, which is the one use where I do not mind renting. No source to fork, no key of mine, no offline. Scriptable, not ownable.