Engineering article reviewResearch Insights Made Simple
Long-running agents: shift handoffs
Carrying a project across context windows: environment, feature list, progress log and checks
/ Research Insights Made Simple · Agent harnesses 2/7 · Shift handoffs
Slide contents
1. Long-running agents: shift handoffs
Carrying a project across context windows: environment, feature list, progress log and checks
2. Every session starts without memory
3. Compacted conversation is not working state
4. An oversized step breaks off halfway
5. Visible progress does not prove completion
6. The first session sets up the rest
7. Four artifacts answer four questions
8. Features are written as observable scenarios
9. Status changes only after behaviour is verified
10. Models overwrite JSON less often than Markdown
11. Check the foundation before building further
12. One feature gives a verifiable step boundary
13. API success leaves the UI unverified
14. Verification depends on what tools expose
15. Commits preserve recovery points
16. A useful record ties results to evidence
17. Both roles answer every failure
18. Engineering experience supplies measurable hypotheses
19. Test the harness with forced context resets
20. Long-running work is built from verifiable steps