Almost all the coding agents are unable to work more than a few mins without needing a human in the loop. You pretty much babysit them even in an agentic mode.

Add in long context codebases, dependency graphs and you do more work being frustated with their output rather than enabled with it. The first 80% of boilerplate - is like a template they absolutely steam through - but work on the 20% that matters and they fail miserably.

Things that have worked to bring in better context - Project spec plans - Architecture for the whole project - Detailed design breakdown for the modules - TODO for tracking implementation - and a scratch implementation doc

This is the hierarchy of representations that brings an agent upto speed even as it muscles through the individual bits.

How do you deal with long-chunk codebases - and what's been the longest / most detailed project you have used an agent on?