Disclosure: I build Grunz, a coding agent, so I’m answering from that side rather than neutrally.
The thing that surprised me most is that shared context between agents is usually the wrong default. It’s the intuitive design and it fails fast, because every agent inherits every other agent’s transcript and the window is gone before anyone does useful work. Most hosted open models serve 32K and only a handful reach 256K, so the budget is tighter than people expect.
What has worked better for me is passing artifacts instead of transcripts. Each agent gets a task spec plus the specific files it needs, does its work, and writes results back to disk. The filesystem ends up being the shared memory, and it has the nice property of being inspectable when something goes wrong. Handoffs become “here is the path to what I produced” rather than “here is everything I thought about.”
On task ownership, the concurrency bugs I’ve hit were almost all two agents editing the same file. A single-writer rule per artifact removed most of them, and it is a much cheaper fix than a locking scheme.
The unglamorous conclusion is that compaction quality matters more than orchestration cleverness. Once the context ceiling is the binding constraint, how well you summarise state between turns decides whether the system works, and no amount of coordination design compensates for that.
Curious what you’re seeing with Sharkly on the handoff side specifically, since that’s the part I still think is unsolved.