#4807 is the export worker running out of memory on files over 200 MB. I
linked it to the open engineering issue and replied with the workaround.
Triage #4821 and draft the reply. Use the refunds playbook.
›Readplaybooks/refunds.md0.4s
›Memorysearch "duplicate charge"1.2s
Paid twice on 12 Jul. The second charge was refunded on 15 Jul, but the
confirmation bounced: the address on file has a typo. Drafting a reply with
the refund reference and asking them to confirm the address.
A workspace: where an agent runs, and where you set it up and work with it.
Whichever core you already run
Claude Code, Codex and Goose run here the way they run on your machine. Switch between
them later and the agent stays as it is.
Claude Code
Codex
Goose
Others
01 / Build
An agent is a prompt, its skills and its tools, on a frontier core.
Tune it once, and the team has it
Those three pieces live in a shared library with scopes and versions. Point a workspace at
one and it follows every update, instead of five copies drifting apart.
Library6 workspaces reference these
Prompts
Skills
MCP
Commands
Templates
support-triageteamv5
refund-replyteamv2
escalation-rulesprivatev1
tone-of-voicepublicv7
New promptImport from workspace
Agent middleware: the hard parts already run here
Middleware is the layer beneath the agent: capabilities the platform runs, so every
workspace has them from minute one.
AGENT
render the promo cut
ffmpeg
check the signup flow
playwright
summarise last week
pandas
none of these live in the agent’s own container
jrottenberg/ffmpeg
$ffmpeg -i raw.mov promo.mp4
done, and thrown away
mcr.microsoft.com/playwright
$npx playwright test signup
python:3.12-bookworm
$python tickets.py --week
AGENT
sign in and export
browser
$agent-browser open
$agent-browser snapshot -i
$agent-browser click
app.example.com/reports
this bit needs a person
you take the keyboard, same session
@agent/researcher compare the two vendors
coordinator
search · docs
search · news
search · pricing
check · claims
check · numbers
check · sources
writer
promptfind pricing, docs and recent news
result14 sources, 3 of them conflict
promptverify the three conflicts
result2 of 9 claims dropped
WORKSPACE A
draft the notes
read_write
WORKSPACE B
add the pricing
read_write
WORKSPACE C
summarise both
read_only
/mnt/afs/notes
drafts
vendor-a.md
2.1 KB
pricing.csv
12.6 KB
WORKSPACE
note what we settled
another workspace
read_write
read_only
notes
/mnt/memory/k3f9x2ab/
policies
/mnt/memory/7dq2m4xz/
FUSE
store_id
path
updated
k3f9x2ab
/MEMORY.md
2d ago
k3f9x2ab
/project/q3.md
5d ago
just now
7dq2m4xz
/refunds.md
1w ago
AGENT
url/_cp/mcp/…
no token in here
CONTROL PLANE
attaches your token
Bearer ••••••••
UPSTREAM
arrives signed
Or let it build itself
Every piece above is a field you can fill in yourself. Builder mode is the other way round:
you describe the goal, and the agent configures its own core, prompt, skills and schedule.
02 / Distribute
Every entrance leads to the same agent.
Slack threads, HTTP calls, CI webhooks and its own schedule all open one workspace, with the
same prompt, the same skills and the same memory behind them.
Native UIthe console the platform ships
Chat channelsthe thread becomes the session
HTTP APIstream it, or hand off the turn
Webhookfilters run before a session exists
Scheduleeach firing is its own session
Workspace
Sessions
uiwalk me through last night's failures
chatthe deploy is red, what happened?
apiTriage the failing build, open an issue.
webhookBuild api-e2e failed on main, read the log.
scheduleSummarize yesterday's failed builds.
The same workspace, served two ways
Static
1 replica
4 vCPU · 8 GiB, held at all times
Up whether or not anything is asking. It answers the moment a turn arrives, and holds its
machine in between.
Auto-scalingPreview
0 to 5 replicas
1 vCPU · 2 GiB per replica, held while it runs
Replicas follow the traffic: up at once, down once they drain, none at all while it is
idle. Files sit on a shared volume, so the next turn wakes it with them intact.
Both sizes are an example. CPU and memory are set per workspace.
03 / Optimize
The agent proposes. You approve.
Ask it to read its own sessions and it comes back with a change to itself. You approve the
change, and the platform applies the copy you approved.
Agent · builder mode
list_sessions1,240 sessions in this workspace
get_session_export_urls12 transcripts, read in full
workspace_prompt_proposereq_8f2c41 · the id, not the change
workspace_prompt_applyapplied
agent_requestspending
- Reply in the language of the ticket
+ Reply in the language of the customer's last message
the change is held here, not by the agent
ApproveRejectthe owner decides
propose
re-reads what you approved
Workspace · system promptupdated
Deployment
One control plane, as many clusters as you need.
Control plane
The cluster you install on
Workspaces run beside the control plane.
one outbound connection
A cluster behind NAT
Its runner dials the control plane and keeps the line open.
A customer site
Their cluster, their network, your workspaces.
One installer, one node or a cluster
The same install targets a single k3s host or a Kubernetes cluster you already run.
It can also install with no internet at all, from a bundled image set and a registry
it brings up itself.
Environments dial out
A runner inside the cluster opens one connection to the control plane and works over
it. There is no inbound route to open, and no kubeconfig to hand over.
Get started
One command on one machine
$curl -sfL https://docs.neutree.ai/nap/get.sh | sudo sh -
One Linux machine · 8 vCPU / 32 GB / 200 GB · around ten workspaces
