For infrastructure
Will it run, scale and recover in our estate?
The questions a review asks of Promenaut, answered from the platform itself, and everything built for them.
The questions, in the order they come.
Four services, the web interface, the API, the orchestrator and the agent workers, run as containers in private subnets with no public addresses.
- Compute and data sit off the internetServices and data stores run in private subnets with no public addresses.
- Queue-driven service architectureWork moves between the API, orchestrator and workers as queued jobs on Redis rather than direct calls between services, and a job held by a process that stops is picked up again, so a slow or restarting service delays work instead of dropping it.
- Platform event busServices publish progress to Redis channels which the API fans out to connected browsers, so a worker reporting progress does not need to know who is listening.
The full answer is in How Promenaut Runs in Your Estate.
Each service on the signal that matters for it: the API on processor load and open real-time connections, the orchestrator on the depth of the workflow queue, and agent workers on the depth of the agent task queue, which the platform publishes itself.
- Each service scales on its own signalServices scale on the signal that matters for each: the API on processor load and open real-time connections, the orchestrator on processor load and on the depth of the workflow queue and of the platform job queue, and agent workers on the depth of their two task queues together, one for agents that call models and one for deterministic agents, which the platform publishes itself.
- Horizontally scaled workersAgent execution runs in worker services separate from the API, scaled by adding replicas on queue depth rather than by making one process bigger, with a budget per model route that is shared across every worker, so a burst against one provider does not use up the allowance of another.
- Long work does not hold an orchestrator slotWhen a workflow dispatches a team of agents, the orchestrator records the batch, checkpoints its state and releases the queue slot it was holding rather than waiting.
The full answer is in How Promenaut Runs in Your Estate.
The work carries on. Jobs retry with exponential back-off, a job held by a process that stopped is picked up again once its lock lapses, and a second request to start the same work is merged into the one already queued.
- A failed or abandoned job runs againWorkflow and agent jobs retry with exponential back-off, a job held by a process that crashed is re-run once its lock lapses, and a second request to start the same workflow, schedule slot, task or rendition is merged into the one already queued.
- A stalled workflow is noticed and restartedOrchestrators re-queue workflows that are in progress but no longer moving, at start-up and on a heartbeat once a minute, with a claim so two instances do not recover the same workflow, a cap on attempts after which a workflow that is still silent is failed with its reason, and a skip for workflows that are legitimately waiting.
- Workflow checkpointingLong-running workflows checkpoint their state as they advance, so a worker restart resumes from the last checkpoint rather than starting the workflow again.
The full answer is in How Promenaut Runs in Your Estate.
The database, both Redis clusters, shared file storage and the object store are encrypted at rest, and connections to Redis, the database proxy and file storage are encrypted in transit.
- Encrypted at rest and on the way to the dataThe database, both Redis clusters, shared file storage and the object store are encrypted at rest.
- Credentials come from a secrets storeProvider keys, signing secrets, integration tokens and database and Redis credentials are held in AWS Secrets Manager and injected into services when they start, rather than set as plain configuration.
The full answer is in How Promenaut Runs in Your Estate.
One console pairs every running component, from containers and queues to the database, Redis, storage, the load balancer and model providers, with the workflows using it.
- Platform X-ray operator consoleAn operator console pairs each running infrastructure component with the workflows using it, and exposes a fixed set of audited admin actions.
- Infrastructure alarmsAlarms are defined for service processor and memory load, database load, free storage and connections, proxy connection failures, Redis evictions and memory, load balancer errors, and ended workflow runs still missing their model risk verdict, with a dashboard, and each alarm is delivered to a notification topic the environment subscribes its operations contact to.
- The platform notices stalls and backlogs before a person doesScheduled health checks look for workflows that have stopped moving, a growing queue backlog, workflow jobs failing and a lost Redis connection, and raise alerts at warning or critical severity to the operations console in real time.
The full answer is in How Promenaut Runs in Your Estate.
A closed set of named actions against live components: stop, pause or resume a run, retry a message, drain or resume a queue, evict a checkpoint, restart or rescale a service, close a streaming connection.
- A fixed set of operator actions against running infrastructureThe operator console exposes a closed set of named actions against live components: kill a run, pause or resume one, retry a message, pause a queue's intake, clear its waiting jobs or resume it, evict a checkpoint, restart or rescale a service, close a streaming connection, replay held event deliveries, reactivate a suspended agent, set the sandbox runtime new skill runs are given, set the install-wide cap on skill runs, keep a launcher pool's hosts up for a set time.
- Safe operator actions without destructive authorityOperator actions are split in two.
- Closing intake and changing capacity, separatelyIntake to a queue can be closed while the work already accepted continues, and reopened afterwards; deleting the jobs still waiting is a separate action.
The full answer is in How Promenaut Runs in Your Estate.
A release is cut on a short-lived release branch.
- Releases are cut, tested and finished on a release branchA release is cut as a short-lived release branch.
- One feature release at a time, patches from a shipped versionOnly one feature release branch may be open at a time; cutting a second is refused until the first is finished.
- Test environments cannot drift from what they mirrorEach change to environment settings is checked before it merges.
The full answer is in How Promenaut Runs in Your Estate.
A deployment to a named environment is a change to a pin file naming the release, merged by review, followed by a plan anyone can read and an approval before anything is applied.
- A client deploy is a reviewed changeA deployment to a named environment is a change to a pin file naming the release, merged by review and approved before anything is applied.
- Each release is rehearsed through the client pipelineEach stable baked release is deployed end to end through the client's own pipeline script, vendored with a checksum guard, in an isolated account, verified, and then torn down.
- Pipelines hold no cloud keysPipelines reach AWS by exchanging the identity GitHub issues each run for a short-lived role session, rather than holding long-lived access keys, with separate roles for building images and for deploying.
The full answer is in How Promenaut Runs in Your Estate.
Through three separate migration trees, for the core platform, each module and each tenant, so a module or a single tenant can evolve its schema without a platform release.
- Layered migration systemSchema changes live in three separate trees: core platform, per module and per tenant.
- Migration safety checksThe migration lane refuses to finish on an unsafe result.
The full answer is in How Promenaut Runs in Your Estate.
Yes. The platform ships with a kit to bootstrap, deploy, upgrade, verify and roll back in your own account, including a full bundle carrying the images and signed modules for an environment without outbound access.
- A client can run it in its own accountA client can bootstrap, deploy, upgrade, verify and roll back the platform in its own AWS account with a shipped kit, including a full bundle carrying the images and signed modules for an environment without outbound access.
- Per-tenant module profilesEach environment ships a declared set of modules resolved at build time, so two tenants can run different capability sets from one codebase without branching.
- One call, several model providersAgents run on models from Anthropic, OpenAI and Google directly, or on models hosted in Amazon Bedrock, through one call.
The full answer is in How Promenaut Runs in Your Estate.
What the platform does for you, layer by layer.
Home and conversation
- Live workflow updates
- A workflow card shows how the run really ended
- Long lists page on the server
Models and routing
- Per-agent model routing with fallback
- A failed model call is retried by kind of failure
- A model call that overruns is cancelled, not abandoned
- Request settings adjusted to what each model accepts
- Agents stay bound to models that exist
- Move agents from one model to another in one step
- Native tool calling across model providers
Output checks and repair
- Structured output is repaired before it is refused
- A quality controller is built and not yet wired in
- A model-output validator is built and not yet wired in
Guardrails and model limits
- Input too large for one call is split and merged
- Agents can cap what they send
- Model limits tuned without a redeploy
- Guarded calls paced to the account's guardrail quota
- Output length fitted to the model, and truncation reported
- Images sized to the model before they are sent
- Guarded model traffic is measured, and each run logs its limits
Documents and media
- Document pieces sized to the model and guardrail limits
- Documents split along their own structure
- One failing piece does not fail the document
- A failing piece is halved rather than repeated
- Parallel pieces share the model allowance
- An unreadable upload stops the run instead of passing silently
- Screenshots and diagrams are read by a vision model
Registry and versions
- Versioned agent registry
Fleet and control plane
- Per-agent control plane
- Capability bench depth
- Readiness-first agent roster
- Fleet lifecycle attention queue
Teams and workflows
- Rule-triggered service teams
Handoffs and context
- A team whose agents do not actually connect
- An edit proposal is checked before it is used
Execution
- Workflow checkpointing
- Batched team execution with resume
- Fan-out over a derived task list
- Long work does not hold an orchestrator slot
- A re-entered batch recovers instead of hanging
- Checkpoint cost is tunable
- A workflow cannot be marked complete with agents still running
Run control
- A running workflow can be paused, stopped or cancelled
- Workflow replay
- Stopping, pausing and resuming work that is already running
- A run that asked a person waits in a status of its own and resumes on the answer
- A stop cancels the model calls already in flight
- A resumed or restarted run is driven by its newest job
- Safe to retry a workflow start
Scheduling
- Scheduled workflows
- One organisation's schedules do not crowd out another's
Ingestion and provenance
- Unchanged documents are not extracted twice
Identity and access
- Safe operator actions without destructive authority
Approvals, decisions and containment
- Runaway agents are suspended automatically
- A quarantined module's agents do not run anywhere
- An agent that keeps failing on its own configuration is contained automatically
- A control directive can slow matching work down without stopping it
- A control directive can pause, drain or kill one agent or one team inside a running workflow
- Automatic step-up of containment for a runaway agent
Model risk
- Agent run health and trends
- One finding for failures that share a cause across the fleet
Modules and manifests
- One broken module does not stop the platform
Module interface
- A module installed after the build still draws its own interface
Build and release
- SDK packages released in dependency order
Services and data
- Queue-driven service architecture
- Layered migration system
- Migration safety checks
- Platform event bus
- Services reach the database through a pool
- Installed modules sit on read-only storage
- One interface over local and object storage
Estate security
- API calls and sign-in are rate limited
- Request bodies are limited before sign-in is checked
- A skill script runs in the sandbox runtime the install chooses
Scaling and recovery
- Each service scales on its own signal
- A failed or abandoned job runs again
- A stalled workflow is noticed and restarted
- Horizontally scaled workers
- Work that failed everything has somewhere to go
- Live update connections are bounded
- Each model route has its own token and request budget
- A workflow waits for model capacity instead of failing
- Workflows can run on synthetic model responses
- Guardrail limits measured, not assumed
Builds and releases
- Content-addressed image builds
- Candidate and released are separate registries
- A client release with a fixable critical flaw fails
- Releases are cut, tested and finished on a release branch
- One feature release at a time, patches from a shipped version
- Release rulesets applied from code
- A release image says exactly what it was built from
- Profile lock files
- Version bumps derived from commits
- Release notes written from the tickets in a release
- Web first-load chunking with a build-time loop check
Deploying environments
- A client can run it in its own account
- A merge deploys the internal environments
- A client deploy is a reviewed change
- Each release is rehearsed through the client pipeline
- A deploy can be rolled back
- Tearing down an environment needs approval
- A surface can be switched on per environment
- Staging only moves forward unless its database is reset
- Test environments cannot drift from what they mirror
- A deploy checks every artifact before changing anything
- Each install gets the reference data meant for it
- Schema changes land before new code runs
Operations and observability
- A fixed set of operator actions against running infrastructure
- Closing intake and changing capacity, separately
- Platform X-ray operator console
- Tracing and metrics
- Infrastructure alarms
- A service on a different build is flagged
- The platform notices stalls and backlogs before a person does
- What an environment is really running
- An operations console for running work and the agent fleet
Weighing us against the tools you already know? See how we compare, with sources →
The full answers, written for you.
How Promenaut Runs in Your EstateAbout 20 minutes
Promenaut Platform InternalsAbout 60 minutes, in four parts that stand alone
Bring your platform team.
Plan how Promenaut runs, scales and recovers in your own cloud account.
Promenaut
