Coding agent capability has moved fast enough that a lot of the skepticism floating around is a year or two out of date. Here’s an honest current read.
Multi-file features with clear requirements, test generation for code with well-understood behavior, and migrating between frameworks following an established pattern are genuinely reliable now. Multi-agent setups, covered in more depth in our explainer on how they actually work, split work across specialized subagents for testing, docs, and review, running in parallel and finishing faster than one agent working sequentially.
They’re still weak on ambiguous requirements needing real judgment calls, architectural decisions with long-term tradeoffs, and legacy code with undocumented, load-bearing weirdness. An agent can confidently “fix” something that was actually a necessary workaround.
The honest measure isn’t a benchmark, it’s your team’s actual acceptance rate on generated pull requests over real weeks of use. See Terminal-Bench for current independent results, but test on your own codebase before trusting any number.
