// operating system · $15

Route by capability, run bounded cycles, and resume from disk.

A small CommonJS agent kernel with capability seats, a by-kind router, bounded goal cycles, JSONL memory and ledger state, continuity snapshots, budgets, and a deterministic echo adapter.

flux@store
# Setup outline · see included README
› Classify the task kind
› Resolve a capable seat
› Act through echo or your adapter
› Record the cycle and continuity state
agent-os · portable workflow
1Classify the task kindscoped
2Resolve a capable seatscoped
3Act through echo or your adapterscoped
4Record the cycle and continuity statescoped
Workflow overviewRead · configure · verify
$ run --pipeline
1
Classify the task kind
2
Resolve a capable seat
3
Act through echo or your adapter
4
Record the cycle and continuity state
$ explain --how-it-works

How it actually works

Exactly what it does — and when.

1Route task kinds to logical capability seats
2Run bounded classify-route-act-record cycles
3Persist JSONL memory, ledger entries, and a continuity snapshot
In practice · Prototype locally, then connect a model

A developer runs a complete goal through the deterministic echo adapter, inspects the cycle trace and persisted state, then registers a bring-your-own adapter behind a coder seat. The kernel stays the same while the provider implementation changes.

$ cat spec.txt
IncludedKernel, router, store, adapters, budgets, catalog, CLI
SetupRun demo, init config, then add seats
DependenciesZero runtime dependencies
BundleStandalone or in the $199 bundle
$ benchmark --vs-alternatives
This product
A hand-wired agent script
Routing
Capability seats and by-kind lookup
Provider names at each call
Cycles
Bounded classify-route-act-record loop
Custom control flow
State
JSONL ledger and continuity snapshot
Often process memory only
Providers
Echo adapter or bring your own
Real model access is still your responsibility
$ diff --before-after
Without it
  • Provider names are glued into call sites
  • A restart loses mid-flight state
  • A prototype needs paid model access before it can run
With Agent OS
  • Swap models behind logical capability seats without changing call sites
  • Run the full loop locally with the deterministic echo adapter
  • Persist memory, ledger entries, and bounded live state across restarts

Use it to…

Real ways people run Agent OS.

Route work by task kind

Register logical seats for code, summarization, or other capabilities and bind your own adapters.

Prototype an agent loop locally

Run goals through the echo adapter before connecting a paid or remote model.

Resume a bounded workflow

Read the continuity snapshot and JSONL state after restarting the process.

What's inside

Task-kind routing
Bounded cycles
Local records
Zero dependencies
MIT-licensed source

Who it's for

Developers building a small agent runtime

need routing and state without a hosted framework

Teams bringing their own models

want call sites separated from provider names

Builders prototyping bounded cycles

want a keyless deterministic demo first

Questions, answered

✓ Instant delivery✓ Yours to own✓ Use on every project✓ Built to ship results

Ready to put Agent OS to work?

Instant download · one-time price · own forever · all sales final

Agent OS