Question 1

Modal's Sandboxes provide isolated execution environments — why not just build agents on Modal instead of using AgentSky?

Accepted Answer

Modal Sandboxes give you an isolated ephemeral execution container — the same layer E2B provides. You still need to write the agent loop, wire the LLM, handle long-running state across crashes, implement retry logic, and build or bring every connector. AgentSky ships all of that pre-integrated across eight maintained harnesses. Modal Sandboxes are an infrastructure primitive; AgentSky is the assembled agent above it. They can be composed: Modal runs compute-heavy inference, AgentSky orchestrates the agent task.

Question 2

Modal launched a token-metered Shared API in July 2026 — does that now compete directly with AgentSky?

Accepted Answer

Modal's Shared API meters tokens for models served on Modal's infrastructure — it is a model-serving and inference layer. AgentSky is an agent-execution layer: it calls any model (including BYO subscriptions), wraps it in a maintained harness with persistent state and crash recovery, and exposes 2000+ connectors on the same call. The two are complementary: Modal runs the model at low inference cost; AgentSky orchestrates the multi-step agent task that calls it.

Question 3

Can Modal run Claude Code, Hermes, OpenClaw, or the other AgentSky harnesses as cloud agents?

Accepted Answer

You can package any harness into a Modal function or Sandbox if you maintain the image yourself. However, six of AgentSky's eight harnesses — Hermes, OpenClaw, pi, DeepSeek Harness, Kimi Code, and opencode — have no official managed cloud form, meaning you would need to maintain the container image, handle credential rotation, implement state persistence with Modal Volumes, and build crash recovery. AgentSky maintains all eight.

Question 4

I need GPU access for fine-tuning alongside agent tasks — can AgentSky handle the training workload?

Accepted Answer

AgentSky is optimized for long-horizon agentic task completion, not GPU-intensive ML training or batch inference. For fine-tuning a model, running LoRA adapters, or high-throughput parallel inference, Modal's GPU marketplace (T4 through B200) and per-second serverless scaling are the better fit. The two tools address different workload shapes and can be used together: run fine-tuning on Modal, then route the fine-tuned model endpoint into an AgentSky agent.

Question 5

Modal requires Python — does that limit which teams can build agents on it?

Accepted Answer

Modal's SDK is Python-native, so defining Modal workloads requires Python @app.function() code. Inside a Sandbox or function you can run any language or binary. AgentSky requires no language-specific setup — you call a REST API to start an agent and the platform handles the runtime. For teams whose stack is not Python, or who want to delegate tasks without writing infrastructure code, AgentSky has a meaningfully lower entry cost.

	Cloud agent platformAgentSky	Serverless GPU computeModal
What you get	A running cloud agent — harness, model, persistent state, and channels — behind one API call	Serverless compute primitives (functions, Sandboxes, Volumes, Queues) that you wire into agent behavior in Python
Primary use case	Delegating multi-step agentic tasks to maintained, crash-resistant harnesses with built-in channel integrations	ML training, batch inference, large-scale data processing, and GPU-intensive serverless workloads
Time to first agent run	Seconds — pick harness and model, call the API	Hours to days — write @app.function() agent logic, wire the LLM, manage Sandbox lifecycle, implement crash recovery
State and persistence	Persistent computer; survives crashes, network loss, and laptop closure; resumes mid-task automatically	Modal Volumes provide durable storage; Sandboxes are stateless by default; crash recovery is your application code's responsibility
Model choice	8 harnesses x any supported model; BYO Claude or ChatGPT subscription reduces model cost to $0	Model-agnostic for custom inference; own token-metered Shared API launched July 2026; bring any model at API rates
Channels and connectors	2000+ built-in connectors (Gmail, Sheets, Slack, GitHub, Linear, and more) declared as add-ons on the same API call	No built-in connectors; integrations are Python functions you deploy on Modal compute
Pricing model	$3 free credit, no card; pay per task at list pricing; BYO subscription cuts model cost to $0	Starter: free with $30/month credits; Team: $250/month; GPU from $1.10/hr (A10G) to $4.29/hr (H100); CPU $0.0000131/core/sec
Best for	Running agents immediately without building or operating agent infrastructure	GPU-intensive ML workloads, custom inference servers, fine-tuning, and serverless compute for non-agent use cases

AgentSky vs Modal

How they compare

Modal is the right choice for these workloads

Common questions

The agent and its computer. One API.