Agent sandboxes commoditise, and the value moves to saved state
Runtimes that save, branch and restore an agent's workspace, paid for by teams running long coding and ops agents.
Crowded · $0.37 sandboxes, snapshot search gains
Written for: Investor or analyst
Next step
Thesis: sandbox compute becomes a commodity, and margin moves to whoever manages saved state and escalation across long agent runs. Confirming signal: a sandbox provider prices snapshots and branching separately from compute and reports that line growing faster. Already building: Agent37 on price, Daytona on snapshots, and Anthropic in its own orchestration layer.
Why now
Agent37 (YC F26) sells agent sandboxes at under 10% of typical provider cost, with customers averaging $0.37 a month per sandbox, so raw sandboxes are becoming a commodity. A study using Go-Explore-style search over full Daytona snapshots finds that solve rates depend on keeping intermediate progress. Microsoft's Fork-and-Flush clears chat history but keeps the workspace, and gets up to 66% better results on the same compute. Vals AI finds agent teams cost up to 5.1x more for little quality gain. Budgets move from more agents to smarter reuse of one agent's state.
What became possible
Research shows the way to improve agents on a fixed budget is branching from saved workspace states and clearing context, not adding more agents.
Two ways in
- Snapshot-and-branch runtime
A sandbox layer with cheap full-machine snapshots and an API to fork, compare and restore agent runs.
- Customer
- Platform teams running coding agents in CI and on internal repos
- Model
- Usage-based pricing on snapshot storage and branch compute
- Escalation router
Routes an agent's hard steps to a stronger model only when needed, following the ARIA executor-advisor approach, with spend reports.
- Customer
- Engineering leaders watching agent token bills
- Model
- Percentage of measured token savings
Competition
The escalation router faces Anthropic's own advisor tool, which is free to adopt and in-model, plus Not Diamond, which already routes model and effort per step for coding agents. The one open spot is the ARIA idea itself. Today the executor decides when to escalate based on its own judgement, while ARIA uses a learned, cost-penalised controller that works across vendors. That paper is days old and nobody seems to have productised it. Pairing it with per-team spend attribution could be a short-lived wedge. Still, Anthropic or Not Diamond could absorb it quickly, and the vendor-agnostic angle matters most for teams that mix model providers.
Full competition notes (16 competitors)
Both product ideas are already crowded. Snapshot-and-branch is now standard in sandbox products. Morph's Infinibranch sells sub-second full-VM forking. LangSmith Sandboxes reached general availability with forks and tracing. Vercel, Blaxel and Daytona ship snapshot and restore, and open-source layers such as TanStack, Fabric and Helix already wrap fork across providers. So the thesis is real, but the incumbents are the ones acting on it. The remaining room is narrow. Most products capture only the filesystem, not running processes or memory. Few offer a real diff or compare across sibling runs that ties workspace state to the agent trace and the test results. Self-hosted or in-VPC snapshot runtimes for platform teams on internal repos are also scarce, especially now that Daytona has reportedly gone closed source. A new entrant would have to win on one of those, for example running inside customer CI with git-aware workspace diffs, rather than on snapshots alone. The escalation router faces Anthropic's own advisor tool, which is free to adopt and in-model, plus Not Diamond, which already routes model and effort per step for coding agents. The one open spot is the ARIA idea itself. Today the executor decides when to escalate based on its own judgement, while ARIA uses a learned, cost-penalised controller that works across vendors. That paper is days old and nobody seems to have productised it. Pairing it with per-team spend attribution could be a short-lived wedge. Still, Anthropic or Not Diamond could absorb it quickly, and the vendor-agnostic angle matters most for teams that mix model providers.
- Morph Cloud (Infinibranch)
Full-VM snapshot, branch and restore that Morph says takes under 250ms, with unlimited parallel branches, aimed at agents, RL environments and test-time scaling. This is almost exactly the snapshot-and-branch runtime idea. The speed figures come from Morph's own marketing.
- Morph fork-VM for AI agents
Morph's own page pitching 'branch a run instead of restarting it' to coding-agent builders.
- LangSmith Sandboxes (LangChain)
Generally available. Offers kernel-isolated microVMs with snapshots and parallel forks, and is pitched on the claim that 10 branches cost about as much as one sandbox. Comes bundled with LangSmith tracing, which covers much of the 'compare runs' piece.
- Vercel Sandbox snapshots
Snapshot an ID, then create from that ID. Captures the filesystem only, not running processes or memory.
- Blaxel
Agent sandbox platform with a REST API to restore a sandbox from a snapshot.
- Daytona
Agent sandbox platform. One comparison lists mid-execution snapshots and forking, and Helix and MintMCP report it went closed source in June 2026.
- E2B / Modal / Runloop
Large sandbox providers adding pause/resume and persistent state. This is the commoditising base layer the idea assumes.
- TanStack AI portable snapshots
Open-source portable snapshot plus a fork_sandbox_snapshot call that works across sandbox providers.
- Fabric harness
Harness-level session.fork() that runs over multiple sandbox backends.
- Helix Snapshotter module
Open-source capture/restore/branch interface over several providers.
- Anthropic advisor tool
Beta since April 2026. A Sonnet or Haiku executor consults an Opus advisor inside one API call, with max_uses caps. This is the first-party version of the escalation router, though the executor itself decides when to escalate.
- ARIA paper (Ask Your Advisor)
Posted about four days ago. A learned meta-controller decides when to consult the advisor, and the authors report recovering 44-81% of the success gap at under half the advisor's cost. No product seems to implement it yet.
- Not Diamond
Model router with a coding-agent product that picks the model and reasoning effort at each step, works with Claude Code and other harnesses, and claims 20-65% savings (vendor figures).
- Martian
Hosted LLM router claiming large cost reductions.
- RouteLLM
Open-source strong/weak router framework.
- advisor-strategy-skill
Open-source skill that spawns a stronger advisor model for Claude Code, Cursor, Codex and similar tools.
Checked 2026-10-13
Risks
- Crowded: Daytona, E2B and the clouds all sell sandboxes and can add snapshots
- Labs ship their own harnesses, such as Anthropic's dynamic workflows, that may absorb this layer
- Snapshot research gains come from benchmarks and may shrink on real repos
- Margins fall as sandbox prices race toward zero
Best on tech readiness and team fit; open space is the weak point (4).
tech 7 · pull 6 · space 4 · fit 7 · build 6
Exposed
Companies that could benefit or be threatened. News and context, not investment advice.
Momentum of its main topic
Sources
- Launch HN: Agent37 (YC F26) – Low-cost agent sandboxes for running agentsHacker News
- Returning to Development States: When Full Sandbox Snapshots Help Coding AgentsarXiv
- Clearing chat history helps AI agents find better solutionsThe Decoder
- AI agent teams waste massive tokens for barely measurable quality gains, research findsThe Decoder
- Ask Your Advisor: Learning When to Consult Stronger Models in Agentic SystemsarXiv
- How Anthropic’s Dynamic Workflows Coordinate Claude AgentsLatent Space
What happened next
- +7d · 20 Oct 2026 ✓ launched Perplexity: Perplexity open-sources SPACE, a VM sandbox with snapshots for agents (15 Oct 2026)
Status history
- 2026-10-13 Flagged as buildable
- 2026-10-15 A product launched: Perplexity: Perplexity open-sources SPACE, a VM sandbox with snapshots for agents