01 · threadone relationship
What is happening
Recent messages, summaries, the current run, and the authoritative journal for this conversation.
rebuilt from the recordAgents
While nobody watches.
A durable principal with memory, skills, and a governed path to act.
Journal — agent:support · thread client-acme/refunds replayed from the journal
14:02:11RunStartedagent:support
14:02:11SkillSnapshotrefund-resolution@3 · pinned for this thread
14:02:12ModelCallfirst attempt failed — wire retry, then sibling fallback
14:02:13ModelCallcompleted haiku-4.5 · $0.0053 · attempt trail journaled
14:02:14ToolCalllookup_order → ok
14:02:19GatePausedzendesk.refund $184.00 → human
14:06:31GateResolvedapproved · user:ryan
14:06:32ToolCallrefund → ok
14:06:35MemoryStoredclient preference · confirm refunds by email
14:06:40RunCompleted$0.0071 · paused time $0.00
The unit of work
Not a prompt. A principal with a job.
A Noorle agent is a first-class actor: named, versioned, and granted only the capabilities it needs. Its model policy, memory, skills, autonomy, and interfaces travel together as one operational object.
Change the model without changing the agent. Update its skills without shifting an active thread. Revoke a capability without rewriting its instructions.
agent:support spec · 7c9e31
identity principal: support every call attributable
models support-balanced retry · route · fallback
state threads + curated memory durable across runs
skills support/* manifest pinned per thread
grants orders · zendesk · files no ambient access
autonomy supervised reads flow · actions ask
surfaces web · email · A2A one identity everywhere
one objectversioned behaviorgoverned effects
Durability, defined
The thread is the durable actor.
A run is one activation of that actor. The journal is its durable event log. If a stream breaks, a worker moves, or a gate waits on a human, the process can disappear without taking the work with it.
thread · client-acme/refunds run · 01J6Z7 · platform-completed
01 · admit committed
Freeze the run
Principal, surface, model policy, and skill references land with RunStarted.
02 · work journaled
Commit boundaries
Messages and effect intents become durable before anything reaches the outside world.
03 · interrupt worker gone
Lose the process
A deploy, broken stream, provider failure, or approval wait can end the local execution.
04 · continue claimed
Replay, then resume
An eligible run continues from a safe boundary; a parked gate resumes with the same run ID.
05 · finish terminal
Close once
One terminal record carries the outcome, attempt trail, and attributable cost.
durable record on every run generation-fenced recovery $0 while parked
The state that lasts
It remembers more than the window.
Conversation state, durable facts, reusable judgment, and working artifacts have different jobs — so Noorle keeps them in different layers.
02 · memoryacross threads
What is worth keeping
Curated facts the agent can explicitly store, recall semantically, update, or forget.
controlled I/O, not transcript sprawl03 · skillsversioned behavior
How to do it well
A compact manifest in context; full procedures and scripts load only when the work calls for them.
snapshot pinned per thread04 · workspacedurable files
What the work produced
One filesystem mounted in code runners, sandboxes, Computers, and plugins wherever the agent computes.
artifacts survive the sessionGranted action
It can do only what you grant.
Agents bind capabilities directly — built-ins, your sandboxed code, and connectors. Every side effect crosses the same admission point before it executes.
granted capabilities
builtin:filesread · write
plugin:[email protected]your WASM
connector:zendeskOAuth brokered
admission · every call
lookup_orderallow
refund · $184ask
refund · $2,400deny
journal witness
- actor
- agent:support
- resolver
- user:ryan
- effect
- zendesk.refund
- cost
- $0.0000
01
Read-only
Observe and reason. Side effects do not execute.
02
Supervised default
Reads flow. Actions pause for a human.
03
Full
Eligible actions flow; hardline and parameter rules still hold.
Autonomy is set per agent and can narrow per thread. The platform enforces it — not the prompt. Read the trust model →
The skills flywheel
The runtime commoditizes.The judgment compounds.
Every run under the gates produces skills — encoded judgment that loads into the next run. Better work, more runs, more judgment, across every client. Anyone can copy the harness; they start the wheel at zero.
noorle/agent-os skills flywheel · turning
127 skills encoded judgment +0 this session
01 · run
Agents do the work
Real client work, every effect crossing the gate.
02 · encode
Runs become skills
Each run distills into a skill — authored, or learned from the journal.
03 · improve
The next run is better
Skills load on demand. Faster, surer, fewer approvals.
04 · compound
Judgment accrues
Better outcomes → more runs → more judgment, across every client.
harnessopen source · commoditizing judgment: yours, compounding
One runtime
One agent. Every entrance.
Surfaces admit work into the same thread actor. Identity, policy, memory, journal, and meter do not change with the client.
01 · product
Your interface
Web, AG-UI, or your own streaming client.
02 · messaging
Where people already work
Slack, Telegram, SMS, and email.
03 · agents
Agent to agent
A2A discovery, messages, tasks, and delegated work.
agent:support
principal
thread
policy
memory
journal
meter
Choose the right surface
Judgment, or a published process?
Agent
WorkflowThe next step depends on judgment.
Open-ended work, ongoing threads, and an actor that carries context and learned procedure forward.
relationship persists · path adaptsThe process should be published.
Named, versioned work with schemas, durable waits, and a run someone else can inspect or operate.
version pins · automation starts it →Jobs, not demos
Give it work with a boundary.
01 · customer support1,000 conversations / mo
Resolve the ordinary. Ask on the consequential.
Answer in Slack and email, remember the customer, look up orders, and refund below a threshold. Larger refunds park for approval.
See pricing and rates →02 · research desk200 deep runs / mo
Research until the brief is defensible.
Search deeply, read sources on a persistent Computer, and deliver a cited brief into the shared workspace. The agent can be reached on demand over its endpoint.
See pricing and rates →