Persistent Machines of Loving Grace
An always-on AI agent runtime for long-horizon or recurring work with evolving skills and goal-driven autonomy.
Bring your own model keys.
Durable execution
A skill keeps running for months.
A task's scope is fixed before it starts: its tools, its instructions, the context it reads. Run nine is as lean as run one, and it begins from what the eight before it recorded.
MA
Maria
Push the spaceship-titanic model past 0.85 accuracy. Start from last week's feature set and keep going until it clears.
Helmsman
Running tabular-model-tuning as a campaign on lbValue, target ≥ 0.8500. This is run #9 of the twice-daily schedule, and I've pulled the four learnings the earlier runs recorded.
Skill|Tabular model tuning · run #9
Engineer featuresRunning · 11 steps
Cross-validating candidate encodings…
memory.store.queryprior learnings · tabular classification4 results · 0.3s
compute.sandbox.execfeature_eng.py --fold 0…
Train baselineQueued
Train variantQueued
Approve submissionQueued
Submit entryQueued
Record learningsQueued
Representative run · a fixed example, drawn with the product's own surfaces.
Conversation
Ask in the space. The Helmsman selects the skill, fills its inputs and starts the run.
human.chat.askSchedule
Cron, one-shot, or on-completion of another run. Leased and de-duplicated, so a fire happens once.
agent.schedule.* · 0 7 * * 1Webhook
A public HMAC-verified endpoint per space. Filter and map the payload, then start a session from it.
POST /v1/webhooks/ingest/…Skills & contracts
A skill is a typed graph of tasks.
Agent tasks, deterministic operations and human decisions are the node kinds. One shape can be scored against another, and the Helmsman rewrites the graph from what the runs showed.
Opstart
Load and profile data
compute.sandbox.exec
Agent
Engineer features
cybernetic-runner
glm-proretry 2
Op
Train baseline
compute.sandbox.exec
Op
Train variant
compute.sandbox.exec
Agent
Tune hyperparameters
cybernetic-runner
claude-sonnet-5
Human
Approve submission
Awaits operator input
Op
Submit entry
api.http.call
Op
Record learnings
learner.record
Fixed agent roles
Every skill runs on the same three roles.
What varies is the skill, the memory and the tools each task is given.
Helmsmanagent
The Helmsman steers. It runs the conversation, and what it attends to shifts with the state of the space. It starts skills and follows up on them, reads accumulated learnings, connects integrations, and writes or adjusts skills.
Task execution is delegated to a Runner.
Runneragent
The Runner executes. Its scope is fixed before it starts: one task, the tools that task needs, and instructions written for it. It completes the task or fails it, and a blocked Runner returns to the Helmsman.
A Runner calls no other agent.
Coachagent
The Coach runs the hill-climbing loop. It reads finished runs as evidence — their scores, their failures and their costs — and proposes a change to the skill.
Applying a proposal requires an operator.
Multiplayer AI
People and agents share one session.
Both act on the same state through the same validated actions, and Role-based access controls (RBAC).
MAMariaviewing run #9
ZHZhutyping…
JDJonaslast seen 14:02
Approve submissionunassigned · anyone with accesswaiting
Chessactivev12you · whiteagent · black
♜♞♝♛♚♝♜♟♟♟♟♟♟♞♞♟♟♟♟♞♞♟♟♟♟♟♟♜♝♛♚♝♜
waiting for a move
The learning loop
Every run improves the next.
A campaign scores every run against a target. The Coach reads the finished runs as evidence — their scores, their failures, the learnings they recorded — and proposes the next change to the skill, which an operator ratifies.
DesignPerformanceRuns
activelbValue ≥ 0.8500 · maximize8 scored runs
targetScore: 0.85seed: 7folds: 5
Best0.8241
Latest0.8241
Target0.8500
To go0.0259
Rolling baselineoverall 0.83 · n=12healthybreach if overall < 0.62
Coach2Reviews finished skill runs and proposes improvements — ratify or reject.
Active reviewreviewingCoach is reviewing a completed run.Watch Coach
Proposals2
Tighten train taskproposedworkflow refinementmediumpreview ok12m ago
Train task timed out twice; reduce estimator count.
update_task_goalhas reflection evidence
Add eval criterionproposedeval changehighpreview ok1h ago
The suite scores accuracy and nothing else.
eval.criterion.add
Integrations & MCP
Integrate with any system, private APIs and MCP tools.
165 operations ship with the platform, and every connector or MCP server a space installs adds more.
GitHub
Slack
Notion
Linear
Jira
Stripe
Twilio
Airtable
Resend
Brave Search
arXiv
PubMed
Semantic Scholar
Wikipedia
GMGmail
CAGoogle Calendar
SHGoogle Sheets
VCVercel
SESentry
CFCloudflare
FCFirecrawl
PGPolygon.io
NANewsAPI
FRFRED
WBWorld Bank
OWOpenWeather
GrantedA task holds only the operations its capability grant names2 of 165 on this task
Coding pipelines
A coding harness runs as one task in the graph.
Claude Code and OpenCode take resolved inputs from the graph and return outputs the platform validates, so coding lanes fan out, converge and stop for a person like any other task.
triage3 issues selected · fan-out
#482null guard in the loaderclaudetest
#486target encoding driftsclaudepatch
inrepoBinding aflowai/modelling · mainissue triage.issues[1]maxTurns 40
bashruff check pipelines/0 errors · 4.2s
readpipelines/feature_eng.py…
outpatchRef —headSha —testsPassed —
#491flaky encoding testopencodeedit
reviewagentapprovehumanpushcode.repo.push
Isolationno docker socket · no host mounts · non-root · cap-drop ALL · https-only remotes
Request access.
Access is limited to a few invited testers, on your own model keys. If you're working on similar problems, or want the architecture in more detail than a landing page can carry, get in touch.