Full thing here: https://t.co/q2kl3xVTkD. The part most explainers skip β about the tool surface, not the loop β is in the middle. Curious where you'd push back?
Open any "agents" piece from the last six months and you'll find a different definition each time. There's one version that holds up across all of them β and it's thinner than the marketing suggests. Full explainer: https://t.co/q2kl3xVTkD
Failure modes are predictable enough to name. Planning errors, tool errors, efficiency errors, stopping-condition errors, prompt injection. If a vendor can't tell you which of these their system fails on, they probably haven't run it long enough.
4/ Audit one overnight run end-to-end. Don't grade individual diffs. Grade whether you can review the whole output in 30 minutes. If you can't, the multiplier doesn't help β it just hides where things broke.
Kanbots claims you can spin up parallel coding agents on every Kanban card and let them run. Strong claim β and one worth pulling apart.
https://t.co/7SZtKi7RCn
3/ Probe Autopilot's subtask-splitting.
The site says it splits parent issues into subtasks across 4 parallel slots. Hand it a fuzzy ticket β the kind a human PM would push back on β and read whether the splits are sensible or just confidently wrong.