Reference
This Stripe example is not about Nexus itself. It is a useful signal that serious coding-agent teams already operate at a scale where context handoff stops being optional.
“responsible for more than a thousand pull requests merged each week.”
Stripe Engineering Blog: Minions + Stripe's one-shot, end-to-end coding agents: Part 2
The actual bottleneck
Teams usually think model quality is the bottleneck. In practice, broken context transfer is often worse: same architecture decisions get rediscovered, same mistakes reappear, and handoffs degrade under time pressure.
Where teams lose time
Session resets
Each new chat starts cold, so engineers re-explain goals, constraints, and prior choices.
Tool switching
Moving from one agent to another breaks flow when past decisions are not carried over.
Static prompt files
Even good `agent.md` or `claude.md` files get ignored in long contexts without active recall.
How Nexus addresses it
- 1. Persist decisions and work logs as memory atoms.
- 2. Recall only relevant context for the current task, not entire history.
- 3. Share one memory pool across agents via MCP and API keys.
- 4. Keep policy and usage visible with plan-level guardrails (limits, 402 behavior, billing).
Bottom line
Better context transfer yields better execution. For teams running multiple agents and frequent handoffs, memory continuity becomes a production requirement, not a convenience feature.