You are here. See how this question connects to other ideas.
Select a node to open its page · Expand to read within the map
One question
Why does an agent need an execution environment?
A model proposes operations; the environment executes them and holds state.
“Save the report” does not create a file by itself. Tools execute in an environment that handles state, failures and resource access. Persistent tasks also face suspension, recovery, logs and lifecycle concerns.
A sandbox addresses some isolation risks; a durable workspace preserves state. Neither automatically establishes business authorization.
ARC is ArcBlock’s runtime, AFS its resource surface, and AOS its operating-system semantics layer. The agent-computer article also describes architectural goals for complete delegation, migration and verifiable history; these must not all be presented as finished capabilities.
Inspect a task
Record separately what the model proposed, what tools executed and what the system retained. The latter two should remain inspectable when models change.
Sources and further reading
Check your understanding
Does generating a command mean it has executed?
No. Execution needs authorized tools, a runtime and result verification.
Continue along a learning path
- From AI to agents that actStep 5 of 5