How to reduce Eve’s Vercel Sandbox bill

Last verified: 2026-09-10 (eve@0.52.5 docs). Source record: docs/eve-sources.md in this repository.

Execution-hosting costs here are Vercel Sandbox compute for Eve’s isolated bash environment. They do not include model tokens, AI Gateway, Vercel Functions, Workflows, or third-party APIs — those stay on the bill when you optimize sandboxes.

1. Idle time after the agent stops working

What to inspect: Sandbox session duration vs. when the model last ran a command. On vercel(), the VM times out after inactivity (default 30 minutes in Eve docs) while the filesystem persists for resume.

What to change: Call sandbox.stop() from authored code when a turn’s shell work is done. Narrow onSession timeouts where your backend supports them. Avoid leaving sandboxes provisioned while waiting on human approval unless you need instant resume.

Tradeoff: Stopping sooner cuts provisioned memory hours but adds resume latency on the next message.

Verify: Compare provisioned GB-hours week-over-week; episode count should stay flat while memory hours drop.

Upstream:

ctx.getSandbox().stop() ends compute without deleting the durable session; Vercel resumes on next I/O.

2. Over-allocated vCPUs and memory

What to inspect: vercel({ resources: { vcpus: N } }) in agent/sandbox.ts. Each vCPU includes 2 GB provisioned memory on Vercel.

What to change: Start at 2 vCPUs; scale up only when builds or tests fail from CPU limits. Match subagent sandboxes to their task size instead of inheriting parent defaults.

Tradeoff: Smaller sandboxes fail on heavy npm install or parallel test runs.

Verify: Active CPU utilization in Usage — if Active CPU ≪ provisioned time × vCPUs, you are paying for idle allocation.

3. Repeated bootstrap and cold templates

What to inspect: Whether bootstrap runs every session. Template reuse requires stable sandbox source, seed files, revalidationKey, and backend options.

What to change: Move one-time installs to bootstrap; use revalidationKey only when external inputs change. Seed agent/sandbox/workspace/ instead of re-cloning large repos each turn.

Tradeoff: Aggressive caching can serve stale dependencies until revalidationKey bumps.

Verify: Creation count vs. episode count — spikes suggest templates not reusing.

4. Subagent and shared sandboxes

What to inspect: Each subagent gets its own sandbox unless explicitly sharing parent.sandbox. Concurrent subagents multiply provisioned memory.

What to change: Share the parent sandbox only when children need the same files and you accept the security boundary. Otherwise cap concurrent subagents in product logic.

Tradeoff: Sharing reduces cost but removes isolation between agents.

Verify: Concurrent sandbox count in the dashboard during peak agent runs.

5. Retries and duplicate episodes

What to inspect: Workflow retries, client double-submits, and channel redeliveries that each resume or create sandboxes.

What to change: Idempotency keys on session creation, dedupe webhook handlers, shorter tool timeouts to fail fast.

Tradeoff: Stricter dedupe can drop legitimate retries on flaky networks.

Verify: Creations/resumes per successful user task (should approach 1:1 for steady-state chats).

6. Snapshots and retained state

What to inspect: Snapshot storage GB-months; default persistence snapshots filesystem on stop.

What to change: Use keepLastSnapshots: { count: 1 } where appropriate; delete abandoned sandboxes; opt out of persistence (persistent: false) for ephemeral CI-style tasks.

Tradeoff: Fewer snapshots mean slower resume or full bootstrap on next run.

Verify: Snapshot storage line item flat or down after retention tuning.

Upstream:

Persistent sandboxes auto-snapshot on stop; snapshot storage is billed separately from compute.

Prioritized checklist

Next steps

• Estimate sandbox cost with your episode and idle-hour assumptions
• Eve agent costs overview for the full diagnose → calculate → migrate path
• Self-host execution when optimized managed sandbox still exceeds operator budget