aflowbeta

Agentic process runtime.

A persistent, multiplayer environment for long-horizon agentic work, built around one conversational orchestrator, reusable skills, stateful applets, semantic memory, and a curated attention layer. A skill is a goal-directed graph with per-task tools and output contracts; memory and recorded learnings can persist across runs. One goal-driven execution model serves different applications: model optimization, portfolio operations, research, and collaborative media production.

Access is limited to a few invited testers — bring-your-own-model-keys.

How it works

Durable execution

A skill runs as a durable, resumable process.

Starting a skill creates a run: a process that survives restarts, resumes from the step that stopped, and records each step's operation, inputs, outputs, duration, token cost and status. A run starts from a conversation, a schedule, or an inbound webhook, and a human-in-the-loop step pauses it for a decision without timing out.

space · modelling / chat

Karim

Push the spaceship-titanic model past 0.85 accuracy. Start from last week's feature set and keep going until it clears.

space · modelling / action center

Needs you3 open

Approve a writehigh risk you

DELETEkaggle.com/api/v1/datasets/versions

{
  "dataset": "spaceship-titanic-features",
  "version": 4,
  "reason": "superseded by target-encoded v5"
}

Reason (optional)

Connect Kaggleconsent

A task needs a Kaggle binding this space doesn't have yet. The run is parked, not failed — it picks up where it stopped once the connection exists.

Ratify a Coach proposalskill change

Raise the fold count from 3 to 5 before scoring. Evidence: fold variance exceeded the score gap on 4 of the last 6 runs.

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.ask

Schedule

Cron, one-shot, or on-completion of another run. Leased and de-duplicated, so a fire happens once.

agent.schedule.* · 0 7 * * 1

Webhook

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 DAG of tasks.

A skill is authored as a graph: agent tasks, deterministic operations and human decisions, with the branches and parallel arms drawn before the run starts. Each task declares what it must return. Two fixed roles execute that graph in every space — the Helmsman steers, the Runner works one scoped task — and the same two serve model optimization, financial analysis and video production, because the domain is declared in the artifacts an operator authors rather than described in prompts.

skill · tabular-model-tuning / designercontract valid

Opstart

Load and profile data

compute.sandbox.exec

Agent

Engineer features

cybernetic-runner

glm-proretry 2

Train baseline

compute.sandbox.exec

Train variant

compute.sandbox.exec

Agent

Tune hyperparameters

cybernetic-runner

claude-sonnet-5

Human

Approve submission

Awaits operator input

Submit entry

api.http.call

Record learnings

learner.record

Approval-gated proposals

Validity is recomputed every time the skill is read, so a proposal that would break it is reported as broken before an operator applies it.

Fixed agent roles

Two roles, many skills.

The Helmsman steers and the Runner executes, and that wiring is identical in every space. What varies is the skill a Runner is handed, the memory it may read and the tools its task grants — which is why one topology serves model optimization, financial analysis and video production. A Coach reviews finished runs and proposes changes an operator applies.

fixed agent roles · identical in every space

Helmsmanagent

The Helmsman steers. It handles the conversation, and what it attends to shifts with the state of the space and the session. It starts skills and follows up on them, discovers tools and operations, 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 that task. It completes the task or fails it, and a blocked Runner returns to the Helmsman.

A Runner calls no other agent.

Coachagent

The Coach reviews finished runs and proposes changes to the skill.

Applying a proposal requires an operator.

Runners fail fast

A Runner that cannot complete its task fails it, which leaves the defect in the skill. Each failure states whether a retry is possible and what to try next.

Multiplayer AI

Several people and the agent work one session together.

A session carries a durable participant roster and live presence: who is here, who is typing, who is looking at which step. Any participant with the required permission resolves the decisions a run is waiting on. Shared state lives in an applet, which declares its state, the actions that may change it, and a view — a person uses the rendered view and the agent receives the same action set as a tool contract.

session · run-09-tuning · 3 participants

KAKarimviewing run #9

MRMiratyping…

JDJonaslast seen 14:02

Approve submissionunassigned · anyone with accesswaiting

ui.applet.act

Chessactivev12you · whiteagent · black

♜♞♝♛♚♝♜♟♟♟♟♟♟♞♞♟♟♟♟♞♞♟♟♟♟♟♟♜♝♛♚♝♜

waiting for a move

two authored primitives

SkillAppletDeclaresA processA shared work objectShapeA typed graph of tasksTyped state and actionsAdvances byTask dependenciesState transitionsProducesA recorded outcomeA state anyone can act onThe agent usesThe task capability grantThe declared action setA personSupplies input and approvalsReads and changes the state

Both read and write the agent's memory, which outlives every run that touches it.

The gateway proves form, not judgement

Every action, from a click or from the agent, is checked against the same RBAC rules, its declared schema and the current version, and repeats are idempotent. Whether the action was a good one is answered by evaluators.

Goal-driven campaigns

Campaigns loop until the goal is met.

A skill declares a goal and the outcomes that satisfy it. A campaign binds it to a metric, a direction and a target, then runs it repeatedly — each run scored, plotted, and recording what it learned for the next run to read. The loop stops when the outcomes are met, or on its run, cost or time budget.

skill · tabular-model-tuning / performance

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

Integrations & MCP

Built-in operations, your APIs, and MCP tools.

A task can call any of 165 built-in operations. Beyond those, registered HTTP API endpoints and MCP servers become callable the same way, installed into a single space and granted to individual tasks.

what a task can call

Orchestrate and delegateworkflow · skill · agent38Reach external systemsapi · mcp · search35Remember and learnmemory · learner23Judge and improveeval · guardrail · proposal22Generate and renderai · ui21Set up a spacespace · catalog · store · integration15Involve peoplehuman · user6Write and run codecode · compute5

GrantedA task holds only the operations its capability grant names2 of 165 on this task

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

Least privilege by default

An agent starts from a small default set, and each task declares the operations it may call.

Models & BYOK

Model choice per task, on your own provider keys.

Frontier models and open weights are addressed through one catalog, with the model set per task. Aflow calls each provider with the credentials you supply.

the model catalog

FrontierGPT-5.6 TerraClaude Opus 4.6Claude Sonnet 5Gemini Pro 3.1Qwen3.7 Plus

Open weightsGLM-5.2Kimi K3DeepSeek V4 ProMiniMax M3GPT-OSS 120B

ProvidersOpenAIAnthropicGoogleFireworksOpenRouter

Coding agent

Design and run full coding and review loops.

Author the coding pipeline, define branches per case, and automate it with human approval gates. Each run records learnings that the next one reads.

code.agent.run · harness claude · repo aflowai/modelling @ mainworking

bashruff check pipelines/…

Isolationno docker socket · no host mounts · non-root · cap-drop ALL · https-only remotes

OutputpatchRef · 1 commit · secret-scannedno push authority

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.