The process dies mid-action
An email, payment or record may already exist when the runner loses the reply. Retrying blindly creates a second effect.
Managed production runtime for agentic workflows
Put agentic workflows into production—without losing control.
Bring a workflow that crosses systems, calls a model or pauses for a person. We design and deploy the production path; mklang platform runs it with managed credentials, recovery and an audit trail.
One email, even when the runner dies mid-send.
workflow contractwho owns what
- your processEvent arrivesmail, webhook, schedule or API
- modelJudgementone bounded mklang step
- your teamHuman decisionpark, reply, resume
- platformExternal effectcredential resolved, effect claimed first
- platformRecorded outcomesame deployment, visible audit
Judgement is a step. Everything else is infrastructure.
Where workflows break
A good decision is not yet a reliable process.
The human handoff lives elsewhere
The run stops, but the decision is buried in chat or somebody’s inbox. Nothing owns the wait or the return.
Nobody can name what ran
Prompts, credentials and deployment state drift until an incident asks for an answer the system cannot reconstruct.
Managed delivery
Bring the process. We build the production path.
This is a managed product, not a workflow editor. We map the process, package and deploy it, connect the systems and onboard the people who operate it.
- Bring
A workflow that matters
Show us the current path, its systems, its human decisions and what must not break.
- Deliver
A controlled production workflow
We implement the judgement, connectors, failure policy, signed deployment and acceptance proof.
- Operate
Your team stays in control
Operators see runs, answer waiting decisions and audit what happened from the hosted console.
Effects are claimed firstRecovery cannot invent a second send.
Waiting is a real stateA person can reply without holding a worker.
Deployments have an identityEvery environment can name the signed artifact it runs.
Failure proof
We test the moment most systems explain away.
The acceptance test kills the runner after the vendor records the effect and before the reply arrives. A second process recovers the run; the ledger settles the ambiguity.
Not a benchmark. A crash between two lines of reality.
$ pytest -k crash --run-real-infra vendor effect recorded, reply withheld runner SIGKILL — mid-effect runner second process recovers the run ledger ambiguous claim → terminal until settled vendor asked again with the same idempotency key PASSED effect is not duplicated PASSED the run finishes when the vendor can deduplicate result one email
Your operator loop
Park, reply, resume.
When judgement needs a person, the run waits in a queue. Reply or abandon it; no worker is held while the decision is yours.
For engineering teams
An open language. A managed runtime.
The model is called for judgement. Triggers, retries, credentials, deployment and external effects remain outside it.
Author in the open. mklang, its specification and the clients that speak the wire are publicly readable.
Run on the hosted platform. The commercial runtime is managed and not published.
$ pipx install mklang
$ mklang console
A machine remains one judgement step inside the workflow. The current connector surface covers Microsoft Graph, Slack and generic HTTP.
Available now
Managed onboarding, scoped to your workflow.
Live production hostRestore rehearsed.
First external workflowNow onboarding.
Delivery + runtimeQuoted after workflow assessment.
Start with the workflow
Show us where judgement meets consequences.
Send the process as it works today: the systems it touches, the decisions a model or person makes, and the cost of getting it wrong. We will assess the production path and scope delivery with you.
Useful context:
- what starts the workflow;
- which systems it reads and changes;
- where a person must decide;
- what failure or duplication costs;
- expected volume and target date.