Augment Code

You will love Augment on the repository that makes other tools flail: the Context Engine holds up on big monorepos where agents usually get lost, and that is the trait that decides it, because finding the right file is most of the work on a legacy codebase and the part an agent gets wrong first. Day to day it means fewer sessions that begin with you pasting paths.

Pick it for the large legacy codebase your team dreads and the tasks that start with a search. Pick Claude Code for a small or new repo, where the index adds a markup and not much else.

The risk is the meter shape. There is no bring-your-own-key, so every token goes through Augment at the provider's list price plus a flat service fee plus compute, and the plan fee is a deposit against that. A long agent run costs whatever it costs, and a run that loops costs that multiplied. Transparent is not the same as capped.

Set a usage ceiling before the first week, because the tool will not set one for you. What it does right: a CLI that runs headless in CI, so at least the runs can be bounded by a script, a timeout, and a budget someone else wrote down.

Two claims, examined separately. 1. SWE-bench Verified, 65.4 percent, March 2025: a Claude Sonnet 3.7 driver with an o1 ensembler selecting among candidates, a pass@k-style result not comparable to single-attempt entries on the public leaderboard. 2. SWE-bench Pro, 51.80 percent, February 2026: Auggie, Cursor and Claude Code run on one identical harness, which is the correct experimental design, and by the post's own account not independently verified.

The second claim is the one to weigh, since a controlled comparison says more than a leaderboard position. What would change the assessment is a third party rerunning the harness. Until then, the numbers are documented and the ranking is reported.

The pricing tells the story: a company that prints its markup on the pricing page is admitting the plan fee is a deposit, not a price, and that gross margin is a fixed percentage of somebody else's bill. That is reseller economics with a proprietary index attached. Exposing the Context Engine over MCP is the other tell, a hedge toward becoming infrastructure that other agents rent, which is a better business than being the agent.

Likely acquirer: an IDE or DevOps platform that wants the index and will discard the chat window. Position: hold; watch whether the fee moves and whether MCP traffic grows faster than seats.

The demo is an agent that finds the right file in a monorepo, which is the demo our engineers actually ask for. Procurement: Business pools usage across up to 50 seats, so sixty engineers means Enterprise, where SSO, OIDC and SCIM live anyway, and a pooled meter means one team's heavy month is everyone's budget. Every plan lists SOC 2 Type II, CMEK, audit trails and exclusion from training, which clears most of the questionnaire before the call.

Onboarding is an extension install and an index build. Approved with conditions: Enterprise contract, a monthly usage ceiling per team, and a report showing who spent the pool.

The auggie repo ships under a proprietary LICENSE.md, so the CLI is a binary I cannot read, hosted on GitHub for the issue tracker. No bring-your-own-key, no local models, and when it misreads my repo I cannot see the prompt that did it. The pattern is familiar: a public repository as a download page.

What I can bend: the Context Engine is exposed as an MCP server, so my own agent, on my own key, borrows Augment's index while I keep the model loop and the logs. That is the piece I would actually use. Grudging respect for shipping the index as a component; none for the binary.