Profiles & Limits#

Rustbox offers two execution profiles, each tuned for different workloads.

Availability: Judge profile is live in the open beta. Agent profile is on the waitlist - and we will invite you when the infrastructure is ready.

Quick comparison#

JudgeAgent
Memory256-512 MB1 GB
Wall time7-15s30s
CPU time4-8s15s
Max PIDs10-102432-512
Networkdisabledfiltered
Best forInterviews, contests, test runnersLLM agents, code interpreters, REPLs
Statuslivewaitlist

Judge mode#

Use case: Competitive programming, coding assessments, interview platforms.

Code runs in a bare sandbox with just the language runtime. No network access, tight memory limits, fast time limits. Designed for evaluating short, deterministic programs against known test cases.

LanguageMemoryCPU TimeWall TimeMax PIDs
Python256 MB4s7s10
C256 MB8s10s8
C++256 MB8s10s8
Java512 MB8s10s1024
JavaScript512 MB8s12s16
TypeScript512 MB8s12s16
Go256 MB8s10s1024
Rust256 MB8s15s64

All languages: network disabled, max code size 64 KB, max stdin size 256 KB.

Typical use: submit a candidate's solution with test case input, compare stdout against expected output. Verdicts (AC, RE, TLE, MLE, PLE, FSE) are determined by kernel evidence, not exit code guessing.

Agent mode · waitlist#

Use case: LLM agent tools, code interpreters, short-lived Python REPLs, notebook cells.

Agent profile is not live yet. - we are opening invites as the Agent infrastructure rolls out.

Higher-memory tier above Judge. Filtered network allows package imports and outbound HTTPS while blocking private IPs and cloud metadata. Sized for typical agent tool-call workloads with room for short-lived Python/JS runs.

ResourceLimit
Memory1 GB
Wall time30 seconds
CPU time15 seconds
Max PIDs32-512 (per language)
NetworkFiltered
Code size64 KB
Stdin size256 KB

Typical use: an LLM agent generates a short Python snippet to parse JSON, call a public API, and return structured output. Round-trip completes well under 30 seconds.

Security#

Both profiles run with identical isolation: 8 kernel layers, seccomp filtering, capability drop, credential isolation, NO_NEW_PRIVS. The only differences are resource limits and network policy.

There is no way to escalate from one profile to the other via the API. Your profile is determined by your API key, not by the request.

Verdicts#

Both profiles return the same verdict set:

VerdictMeaning
ACAccepted - clean exit, code 0
RERuntime Error - non-zero exit or crash
TLETime Limit Exceeded - wall time exceeded
MLEMemory Limit Exceeded - OOM killed by cgroup
PLEProcess Limit Exceeded - fork/thread bomb hit cgroup PID limit
FSEFile Size Exceeded - SIGXFSZ from RLIMIT_FSIZE
SIGSignaled - killed by signal
IEInternal Error - platform infrastructure failure

Every verdict is backed by kernel evidence (cgroup OOM events, signal delivery records, CPU accounting). The platform does not guess verdicts from exit codes.