Globant CODA

The deciding trait is that it is not only a terminal agent. It speaks the protocol editors use to host agents, so the same session you started in a shell can be driven from Zed or from your JetBrains window, and there is a marketplace extension when you want the panel instead. Most terminal tools make you choose once and live with it.

What you will feel is that everything about buying it happens in a conversation. Pick it if you already have a relationship with the vendor. Pick a self-serve tool if you do not.

The safety net has a documented shape and the shape has holes. Checkpoints are built on worktrees, which means an untracked file the agent created, a generated artefact, or anything outside the repository boundary is not part of what gets restored. Run it with the approval flag turned down in a directory that is not a checkout and there is no rollback at all.

What it does right is say so. The dependency on a specific Git version is published rather than discovered, which is more than most vendors document about their own undo.

1. Stating the loop as a sequence rather than describing capabilities is a meaningful editorial choice, because it tells a reader where verification sits relative to the edit, which is the question that determines whether an agent can catch itself. 2. Naming test and lint execution as the tool's purpose puts the check outside the model.

2. What is absent is any measurement of how often the check is invoked or acted upon. The design permits verification; nothing published demonstrates that it happens.

Follow the revenue. A consultancy that sells engineering hours builds an agent to make those hours more productive and to give its salespeople something to demonstrate, so the software does not need standalone margin and will be funded as long as the services business wants it. That is durable, and it also explains the empty price column.

Moat: an existing client list, which is the only moat services firms have ever had. Likely acquirer: none, this stays in-house. Position: low abandonment risk, low pricing transparency, and those come as a pair.

Two things here are genuinely rare and both matter to me. It is documented as running inside scripts and pipelines, which means it can be a measured build step rather than only a thing engineers use, and the macOS binaries are signed and notarised, which removes an argument with my endpoint team before it starts.

What I cannot do is forecast. No published rate means finance gets a range, and a range for sixty seats is not a budget line. Approved with conditions: a fixed quote, a data retention clause, and directory integration confirmed in writing.

The model story is better than the source story. A named local runtime among the supported providers means the weights can sit on my hardware and nothing has to reach a vendor endpoint, and my servers register from a JSON file in the project that I can commit and review like any other config. Hooks and skills extend it without a plugin API.

Then I hit the wall: I cannot read a line of it, and a licence pointer into a README is not a licence I can reason about. Good configuration surface, no ownership underneath it.