Agentsintheloop

Tooling that puts agents on the work around the code rather than in the product: checking branches against their specs, writing and repairing Playwright suites, reviewing pull requests, and reading Figma files directly.

The expensive parts of shipping are not writing the feature. They are proving a branch matches what was specified, keeping an end-to-end suite alive while the UI moves under it, reviewing everything, and getting design out of Figma without paying for an API tier to do it. Each of those is repetitive, each is skipped under pressure, and skipping them is how regressions reach production in a product that handles money.

Each became an MCP server rather than a script, so the same tools are available to any agent session instead of living in one person's terminal. Spec alignment reads the ticket and the test cases and diffs them against the branch. The QA layer drives Playwright across web and API and heals its own selectors rather than failing on a moved button. A Figma plugin paired with an MCP server reads files directly, which is what removes the API tier from the budget. A FastAPI bot puts review and issue-fixing onto the org GitHub, triggered through n8n.

52 MCP tools shipped, 687 QA findings surfaced in a single half-year, and nothing paid to Figma for API access. The tooling is not a demo of agents; it is the reason the review and QA work actually happens on a small team.