There are already four roundups of ephemeral development environments on this site and I wrote all of them. They grade platforms for a human: how fast the editor opens, whether shell history survives, whether you can paste a URL to a reviewer. Those are the right questions when the thing using the environment has hands, a browser, and the patience to click through a consent screen.
A growing share of the environments being created in the world are not created by people, and a coding agent has none of those things. It cannot click "Authorize". It cannot infer from the colour of the text that a build failed. It cannot wait six seconds and still feel good about itself, because waiting is a context window sitting warm and a per-token meter ticking. And critically, it writes the commands: the thing typing rm -rf into your environment is a probabilistic text generator that has read a great many Stack Overflow answers, some of them wrong.
So: six platforms, one rubric, graded for a machine driver. I'm Ajay, I build PandaStack, and it is one of the six, graded by the same rubric including where it loses outright. The human-facing versions are at Top 7 Ephemeral Development Environment Platforms in 2026, Top 8 Ephemeral Development Environment Platforms (2026), Top 9 Ephemeral Development Environment Platforms (2026) and Top 10 Disposable Development Environment Platforms in 2026. The casts overlap this one only partly, on purpose: the agent axis reshuffles the ranking hard enough that four products which look entirely different to a human collapse into a single row here.
The rubric: six questions an agent asks and a human never does
Derive the rubric first and apply it mechanically, or a roundup degenerates into a list of things I happen to like. Ordered roughly by how fast a bad answer disqualifies a platform:
- Is there a real API or SDK that creates an environment in one call — or is the happy path a web UI plus a git push? An agent cannot push to a branch to get compute. Well, it can, but then its sixth hypothesis is a commit in somebody's repository.
- Can the agent read files, write files, run commands and get a structured exit code back? Not a terminal stream — bytes with ANSI escapes, designed for a renderer and a pair of eyes — but an integer plus separate stdout and stderr. An agent reduced to parsing a terminal will eventually decide a build passed because the word "passed" appeared in a filename.
- Branch, fork or snapshot: can the agent checkpoint an expensive state and try six things in parallel from it? Or does exploration mean serial mutation of one box, where hypothesis four is contaminated by one through three?
- Does it cost anything while nothing is happening? An agent that thinks for forty seconds between tool calls is, from a billing perspective, an idle environment generator.
- Isolation boundary: shared host kernel, or its own kernel? The one that matters most when the thing typing the commands is a language model, and the one most agent frameworks never check.
- Teardown: who deletes it, and what does "deleted" mean for the disk? If the answer is "the agent calls delete," ask what happens when the agent crashes mid-task, which is a thing agents do with some enthusiasm.
Notice what is absent: editor quality, extension marketplace, dotfile sync, collaborative cursors, prebuild warm-up. A platform can be world-class at all of it and score zero here, which is not a criticism — it applied for a different job. The fix for question two is boring: on PandaStack it is r = sbx.exec("pytest -q", check=True) giving r.exit_code, r.stdout and r.stderr, with sbx.exec_stream("pytest -q", on_stdout=print, on_stderr=print) as a separate primitive for when a human does want to watch.
The six, graded
1. E2B — built for this, and it shows
An open-source sandbox API aimed squarely at AI code execution: an SDK call in, an isolated microVM out, with filesystem and process surfaces designed for a program to drive rather than a person to type into. The reference implementation of the category.
Question one is an unambiguous yes — creation is a function call, not a deploy. Question two is a yes with the structured-result shape the job needs. For question three, read their current docs on persistence, snapshots and resumption, because that is an area their product has moved in. Question four is the usual sandbox-API answer: you pay while a sandbox exists, so lifetime management is your orchestrator's problem. Question five is the right answer — a VM boundary with its own kernel, which was not the obvious call when they made it. Question six: you delete it, or a timeout does. The honest drawback is that it is deliberately not a human IDE and not an application host, so if the task ends in "and now serve this under production traffic," that is a second product. Being open source answers the acquisition question every infrastructure buyer should ask and most do not.
2. Daytona — re-graded on the agent axis
Included because it scores differently here than in the sibling roundups. In Top 7 Ephemeral Development Environment Platforms in 2026 I graded Daytona as a human-workspace product, which is how it was positioned then. Its current public positioning is explicitly about sandboxes for AI-generated code with an SDK-first surface — a different product answering a different rubric.
Question one is a yes, with an SDK and a declarative environment definition, which is the genuine strength: an agent can be handed a repository and get a reproducible environment without a human first curating a Dockerfile. Question two is a yes on the programmatic-exec shape. Questions three and four go to their docs rather than mine — snapshot or fork semantics and the stopped-state billing story are the two things most likely to have changed since I last looked, and the two that decide whether this fits a branching agent. Question five: verify the isolation model from their security documentation, because "what exactly is the boundary" should be answered in the vendor's own words rather than a competitor's blog post.
3. Modal — function-shaped, which cuts both ways
Python-first serverless compute: decorate a function, declare its image in code, run it remotely including on accelerators, plus a sandbox surface for arbitrary commands. The developer experience is excellent and the cold-start engineering underneath it is serious work.
Question one is a yes if your orchestrator is Python and a more awkward yes if it is not, because the mental model is functions and images rather than machines. Question two is a yes inside the function abstraction. Question three is the interesting one: the primitive is "run this function in this image," so reproducing an expensive state tends to mean re-running whatever produced it rather than branching a live machine — a worse fit for an agent exploring a tree where reaching the interesting state took four minutes of dependency installation. Question four is its real advantage on this axis: the serverless shape means idle is close to free, which matters enormously for an agent that thinks between calls. Question five: read their current security documentation for the runtime boundary, which has evolved over the product's life. The drawback is Python-centricity — a feature right up until the thing you need to run is not Python, at which point you are wrapping shell in a decorator and asking yourself some questions.
4. Vercel Sandbox — ephemeral compute from the deploy platform
A disambiguation first, because the brand collision is confusing. Vercel Preview Deployments are per-branch immutable deploys with a URL and no shell — a reviewer product, graded in Top 9 Ephemeral Development Environment Platforms (2026). Vercel Sandbox is a separate, programmatic code-execution primitive. They share a brand and almost nothing else, so when someone says "we use Vercel for our agent sandboxes," ask which one.
Judging the Sandbox product: question one is a yes, created from an SDK, and question two is a yes on the command-execution shape. Questions three and four — branching and idle billing — go straight to their current documentation, because this is a newer product than most of the field and newer products change fastest. Question five: verify the isolation model from their docs. The ecosystem advantage is legitimate and underrated: if your application already lives there, your agent's environment lives next to it with one fewer vendor, one fewer token and one fewer thing to explain in a security review. The drawback is that the product's centre of gravity is deploying web applications, so heavyweight non-web compute is shopping in the wrong aisle.
5. The devcontainer workspace cohort — Codespaces, Gitpod, Coder, DevPod
Here is the finding that justified writing this post. On the human axis these four are meaningfully different products and I graded them separately, at length, in the sibling roundups. On the agent axis they collapse into one row, because they fail and succeed at the same questions for the same structural reason: they are workspaces for a person, defined by a devcontainer.json or a Terraform template, and the agent is bolted to the side of a product whose centre of gravity is an editor.
Question one is a qualified no — all four can be automated, but creating thousands programmatically is not the shape of the product and the shape leaks everywhere. Question two is a maybe: commands will run, but the primitive you reach through was built to serve an editor's terminal rather than an orchestrator's exit-code check. Question three is a no; you can snapshot an underlying disk if your infrastructure permits, but "branch this running workspace six ways in under a second" is not on offer, because no human has ever wanted it. Question four is where the agent story falls apart: these are optimised for a session, so the most consequential setting in the product is the idle timeout, which on at least one of them lives in a personal preference pane rather than in the repository. Question five is container-on-a-VM for the managed ones and "whatever your Terraform provisions" for the self-hosted ones — the most dangerous thing about that model, because nothing stops you believing you have a VM boundary while running a shared kernel. Question six: a human stops it, and "stopped" frequently still holds a disk.
None of that is a criticism — if a person needs an editor for four hours inside your compliance boundary, this cohort is the correct purchase. It just is not what an agent needs, and the temptation to make it work because your company already pays for one is the most common architectural mistake in this space. You end up driving a browser-shaped product through a CLI-shaped hole, and the first thing to break is the auth refresh.
6. PandaStack — mine, same rubric
Firecracker microVM sandboxes where every create is a snapshot restore. There is no warm pool of idle VMs: the create path allocates a pre-built network namespace, reflinks the rootfs, forks Firecracker and loads a baked snapshot. Create is p50 179 ms, p99 203 ms; the first-ever cold boot of a template, before its snapshot is baked, is about 3 seconds and happens once. Each sandbox gets its own netns, veth pair and tap device from 16,384 pre-allocated /30 subnets per host agent. Guests are Ubuntu 24.04 on a 5.10 kernel under Firecracker v1.16.
Question one: yes — Sandbox.create(template="base", ttl_seconds=900) is the whole thing, with a REST equivalent if you would rather use curl. Question two: yes, plus sbx.filesystem.write(path, contents) and sbx.filesystem.read(path) returning bytes. Question three is the feature I would actually buy this for: sbx.fork_tree(count=4) snapshots the parent once and restores the children from it, so they inherit its memory and disk — you get a VM into an expensive condition once, then explore from exactly that condition at 400 to 750 ms per child same-host and 1.2 to 3.5 seconds cross-host, up to 16 children per call. Question four: hibernate snapshots the sandbox and releases the host, so a resting sandbox costs storage rather than reserved capacity; the rate card is $0.054 per vCPU-hour and $0.0162 per GiB-hour across all classes. Question five: microVM, own kernel, own netns. Question six: sbx.kill(), or the TTL reaper, for when your orchestrator is the thing that died.
Now where it loses. There is no IDE and no web workspace UI for a human to open and poke around in. There is no devcontainer.json support story at all — environments come from a Dockerfile baked into a template snapshot, so if your team's reproducibility contract is a devcontainer file in every repository, PandaStack does not read it and you would be converting.
Three more sharp edges. Firecracker cannot change a guest's vCPU count or RAM at snapshot restore, so guest sizing is baked at template build time; passing memory_mb on a create is not an error, it is silently corrected to the baked value, which is worse than an error because nothing tells you. The guest kernel is 5.10 with one kernel per host, so no per-template swap, and there is no /dev/kvm inside a guest, so no nested virtualisation. And egress is open by default — a few targeted DROP rules exist for known-abuse protocols, but this is not a default-deny network, so if your threat model needs one, that is an allowlist you build in the guest or at your own edge.
One correction, because this is the mistake everyone makes
There are two branching primitives and they are not interchangeable. sbx.fork() reflinks the rootfs and cold-boots the clone, so it does not inherit guest memory: the child gets its own entropy, its own PIDs, and the roughly three-second cold-boot shape. sbx.fork_tree(count=N) snapshots the parent once and restores N children from that snapshot — the sub-second path, and the one that inherits memory and disk. Both cap at 16 children per call. So if the plan is "get this machine into a state where a server is running and a dataset is loaded in RAM, then split it six ways," you want fork_tree; reaching for fork gives you six clean boots of the same disk, three seconds apiece, and a confusing afternoon.
| Platform | Create in one call | Exec with exit codes | Branch from a live state | Idle cost | Isolation boundary |
|---|---|---|---|---|---|
| E2B | Yes, SDK-first | Yes, structured | Verify current docs | Pay while it exists | MicroVM, own kernel |
| Daytona | Yes, SDK-first | Yes | Verify current docs | Verify current docs | Verify current docs |
| Modal | Yes, Python-shaped | Yes, within functions | Re-run, not branch | Serverless-shaped, low | Verify current docs |
| Vercel Sandbox | Yes, SDK-first | Yes | Verify current docs | Verify current docs | Verify current docs |
| Codespaces / Gitpod / Coder / DevPod | Automatable, not the shape | Editor-terminal first | No | Session-optimised; the idle timeout is the setting | Container on a VM, or whatever Terraform makes |
| PandaStack | Yes, 179 ms p50 restore | Yes, exit_code plus check=True | fork_tree(count=N), 400-750 ms same-host | Hibernate releases the host; storage only | MicroVM, own kernel and netns |
The loop, end to end
Here is the rubric as working code: create, write a file, run it, read a structured result, branch from the warm state for a second hypothesis, tear everything down. Awkwardness in any one step compounds across every iteration.
from pandastack import Sandbox
# 1. Create. One call, p50 179 ms. ttl_seconds is the backstop for the case
# where the agent process itself dies holding the only reference.
sbx = Sandbox.create(template="base", ttl_seconds=900,
metadata={"job": "repro-1842", "driver": "agent"})
# 2. Write. The agent produced this text; it has never been reviewed.
sbx.filesystem.write("/work/repro.py",
"import json, pathlib\n"
"result = {'hypothesis': 'off-by-one in the window', 'ok': False}\n"
"pathlib.Path('/work/out.json').write_text(json.dumps(result))\n")
# 3. Exec. An integer and two strings, not a stream to parse. check=False here
# because a failing repro is information, not an exception.
r = sbx.exec("cd /work && python3 repro.py", timeout_seconds=120, check=False)
if r.exit_code != 0:
print("the repro script itself broke:", r.stderr)
# 4. Read the structured result back out of the guest.
out = sbx.filesystem.read("/work/out.json") # -> bytes
# 5. Branch from the LIVE state to test a second hypothesis in parallel.
# fork_tree snapshots the parent and restores the children from it, so they
# inherit its memory and disk: sub-second each, up to 16 per call. Plain
# fork() only reflinks the rootfs and cold-boots (~3 s), losing everything
# resident. Two different tools.
branch_a, branch_b = sbx.fork_tree(count=2)
for branch, patch in ((branch_a, "off_by_one"), (branch_b, "timezone")):
branch.filesystem.write("/work/patch.txt", patch)
res = branch.exec("cd /work && python3 repro.py", check=False)
print(patch, "->", res.exit_code)
# 6. Teardown. Explicit, and the TTL catches what an exception skipped.
for s in (branch_a, branch_b, sbx):
s.kill()Two things there do more work than they look like. ttl_seconds is the only mechanism in the loop that survives the agent process being OOM-killed between steps three and six, which is the failure that actually generates the invoice; the finally block and the careful cleanup code are the happy path, and the happy path is not where leaks come from. The other is that the branches come from a sandbox that is already warm — and the expensive part of an agent task is rarely the hypothesis, it is the clone, the dependency install, the database seed, the build that takes four minutes.
The environment that bills while a model decides what to type
Here is the dark-comedy line item. A reasoning model handed a hard problem may think for a good long while before it emits its next tool call, and throughout that interval the environment it is about to use sits fully provisioned, doing nothing, and on most machine-shaped platforms fully billable. You have built a device that converts model deliberation into compute spend at a one-to-one ratio. Scale it across a fleet and you have, with no malice and considerable engineering skill, invented an expensive way to pay for silence.
The relevant number is not the hourly rate, it is what stays billable in the stopped state — and "stopped" is not one state. Ask specifically: with the environment stopped, am I paying for compute? The root disk? An attached volume? A managed database the task created? A load balancer and a reserved address? I have seen setups where stopping the compute turned off the cheapest of those five, which from a finance perspective is not stopping anything.
Structurally there are three honest answers. The serverless shape, where no machine exists between calls. The snapshot shape, where resting state becomes stored bytes and the machine is released — hibernate on PandaStack snapshots the sandbox and gives the host back. Or you pay for it, which is fine at small volume and ruinous at large, and the transition sneaks up because it tracks agent concurrency rather than headcount. A fourth answer people reach for — "the agent will delete it when it's done" — is a hope delegated to a non-deterministic process. The agent will sometimes not be done, will sometimes decide it is done and be wrong, and will sometimes crash in a way that loses the sandbox identifier, at which point the environment is not idle, it is orphaned. Platform-side TTLs exist because the caller cannot be trusted, for the same reason TCP has timeouts.
The trap: when an ephemeral environment acquires a DNS record
Every platform here describes its environments as ephemeral. Ephemeral is a claim about intent; the invoice is a claim about fact, and the two part company with impressive regularity.
The mechanism is always the same and almost always sympathetic. An agent-driven environment turns out to be useful for something other than the agent — someone notices the sandbox running the integration suite has the best test data in the company. A URL gets pasted into a channel. Somebody bookmarks it, then points a CNAME at it so the bookmark survives a restart, and the moment that DNS record exists the environment has stopped being cattle and become a pet with a hostname. Pets do not get deleted. They get migrated, documented, monitored, and eventually mentioned in an incident review.
The tells arrive in a predictable order: a TTL extended more than twice, a sandbox referred to by name in writing, one that appears in a runbook, one whose metadata a human edited after creation, anything that something other than its creating orchestrator depends on. The defences are unglamorous: an absolute TTL rather than an idle one, a hard ceiling on TTL extensions, no human-editable metadata on agent-created environments, and a rule that anything needing a stable hostname gets promoted to a real deployment with a real owner. If a thing deserves a DNS record it deserves a budget line, and if it does not deserve a budget line it should not have survived the week.
Who did not make the six, and why
- Replit — a strong agent story, but the centre of gravity is a human in a browser-native environment with the agent as a collaborator inside it. Graded on the human axis in Top 8 Ephemeral Development Environment Platforms (2026).
- Fly Machines — a REST API over Firecracker microVMs, strong on questions one, five and six and arguably the cleanest primitive here. It is in the cast of Top 9 Ephemeral Development Environment Platforms (2026), so I left it out rather than repeat the entry. The thing to know: you own the orchestration, so placement, reaping and host-death handling become your code.
- Kubernetes namespace-per-task — common as an agent substrate because the cluster already exists. It fails question five for untrusted code and question four for idle; graded in Top 10 Disposable Development Environment Platforms in 2026. It keeps getting chosen for organisational reasons, which is legitimate but not technical.
- Your own bare metal plus a wrapper script — the honest baseline, and it wins on cost at sufficient scale. Questions one, two and six you answer quickly; three and four are where the real engineering lives.
Picking one
- Code-interpreter shape, model writes Python, you want the best-trodden path: E2B — and read their current persistence docs before designing around it.
- Your orchestrator is already Python and idle cost is your dominant worry: Modal, whose serverless shape makes question four largely disappear. Accept that branching a warm state is not the primitive.
- You want a declarative environment definition so an agent can be handed an arbitrary repository: Daytona, with the isolation model and billing-at-rest verified from their docs first.
- Your application already lives on Vercel and one fewer vendor is worth real money: Vercel Sandbox. Check you are reading the Sandbox docs and not the Preview Deployment docs.
- A human needs an editor for four hours inside your compliance boundary: the Codespaces / Gitpod / Coder / DevPod cohort — the right answer to a question this post is not asking.
- You need to branch an expensive warm state many ways, want idle to cost storage rather than RAM, and want each environment to have its own kernel because a language model is typing: PandaStack, with the baked-sizing, 5.10-kernel, no-IDE and no-devcontainer constraints understood up front.
An agent that thinks for forty seconds between tool calls is, architecturally, an idle environment generator. Price that before you scale it.
Grade on the axis that matches the driver. A platform that is wonderful for a human in an editor is not thereby good for a program creating ten thousand environments a day, and the reverse holds too — I would not onboard a new hire into a sandbox API. The six questions are the durable part; the answers move, mine included, so ask them before the architecture rather than after the invoice.