T1
In-runtime repair
Restart with bounded backoff, then run the runtime doctor when a configuration can be repaired in place.
Performance Faster cores, and an agent that gets back up on its own.
We spent thirty grand benchmarking machines for AI agents. Here is what we measured, and what happens when an agent breaks at 3 a.m.
01Performance benchmark
Cold start
in 1.1 seconds.
We benchmark the runtime itself and keep the setup that gets an agent ready fastest.
Runtime performance lab
Cold-start trial
Mac miniHardware price · $799
5.1sDesktop baseline
MacBook Pro M3 ProHardware price · $3,400
4.0sHigh-end Mac baseline
Molted premiumBenchmarked for agents
1.1sBest measured result
faster than the high-end Mac benchmark.Median of repeated internal trials · OpenClaw cold start · representative hardware
01.1sAfter the benchmark, the operating layer keeps the advantage useful.
SelectionRuntime-tested setup
RegionsEU and North America
Instance isolationPer-instance resource limits
Why not a bigger VPS?
Be aware of regular hosts selling agent hosting.
Their offer is often a standard VPS with more RAM for the price. An agent works one step at a time and each step runs on one core. More RAM lets it run more tasks in parallel but does not make a step faster. A well-built runtime runs well in 3 to 6 GB.
2×faster per core than a standard cloud VM, on the same tasksInternal measurement, median of three runs: 1.7× to 2.1× depending on the task.
Agent steps done in the same minute
Bigger VPSMore RAM, slower cores
5 stepsMoltedLess RAM, faster cores
10 steps2× more- 01
Server processors are slower per core
Budget VPS plans run on server processors built for a very large number of cores. Each core is slower. An agent mostly moves forward on one core at a time.
- 02
Memory scans take processor time
To fit more customers on each machine, budget hosts often scan and merge duplicate memory around the clock. That work uses the same processors as your agent and adds latency.
- 03
We choose processors for speed per core
We test processors on real agent work and keep the fastest per core.
02Built to recover
When an agent fails, Molted repairs it on its own.
Each instance is monitored from inside its runtime and from the outside. When something fails, Molted captures the evidence, escalates through recovery tiers and restores service automatically.
00:00Process stops responding
00:05Heartbeat missed
00:12Post-mortem captured
00:47Instance healthy
Illustrative recovery trace
47sec
2
monitoring layers
4
recovery tiers
Auto
post-mortem capture
24/7
fleet monitoring
Every failed recovery produces a post-mortem and escalates with its diagnostic context intact.
State you can inspect and restore
Every file change leaves a trace.
Inspect what changed. Compare revisions. Restore a known-good workspace, even after a delete. Plain, portable files keep the agent's state independent of its compute.
AutomaticEvery workspace revision is captured
InspectableDiff files between any two points
PortablePlain files, no proprietary memory format
workspace / atlas-prod
savedFiles
▾ workspace
▾ memory/
findings.md
▾ skills/
customer-research.md
SOUL.md
rev 184 · 12s ago
customer-research.mdv183 → v184
# Customer research skill
- Summarize the first result
+ Cross-check primary sources
+ Flag conflicting evidence
Return findings with citations
Illustrative workspace diff · plain files, versioned
Always on, from the first size.
Every size runs on the same cores, from $2.88 a month.