For Builders
01 / 05 — scroll to advance
0.5 Runtime & database for production agents.

Agent infrastructure
for builders.

A3T is a standards-based runtime for agents, paired with a typed database built for agent memory and state. Agents run as Kubernetes pods, wired together on a canvas, backed by memory, tools, and knowledge you define.

A1 → A5 Scroll to see how it fits together
CNCF Serverless Workflow 1.0
Free tier on us
Self-host anytime
Kubernetes ANY CLUSTER · SAME RUNTIME Agent Canvas INPUT / OUTPUT BINDINGS Rt adtech-agent V2.1 · EPHEMERAL Pr payment-rev V1.4 · CHECKPOINT Rb runbook V1.0 · SIGNAL CNCF SWF Agents POD PER AGENT · TYPED RUNTIME RUNTIME BIND HOST
1 Template

Start from an archetype.
Or import your own.

Pick a pre-built template, or import your own workflow definition from any registry. Templates declare the durability contract, variables, and tool surface up front — all as standards-based YAML.

$ a3t template pull ./my-template.yaml

Rt
Ad-tech Agent
Real-time decisioning with post-event learning.
ephemeral v2.1
Pr
Payments Agent
Long-running review with human approval gates.
checkpoint 14d
Tr
Trading Agent
Deterministic replay against historical data.
replay v1.2
+
Your own template
Import any standards-based workflow spec as signed YAML.
standards import
2 Augment

Tools, skills, knowledge, memory.
Agent Database underneath.

Bind an LLM, list the tools the agent can call, declare the knowledge it reads from, and let it write back to memory. Every state transition is queryable AQL against ADB.

All four surfaces share one typed substrate: ADB.

A
adtech-primary
adtech-agent:v2.1 · eu-west-1
4 SURFACES
Tools
ctx-scorer brand-safety tools-registry
Skills
classify_url compute_bid
Knowledge
iab-taxonomy-v3 publisher-kg
Memory
episodic procedural
Storage
ADB · Agent Database
3 Canvas

Wire inputs and outputs.

The Agent Canvas is a Kubernetes CRD you compose visually or in YAML. Bind triggers — HTTP, topics, webhooks, cron — on the input side, emit to webhooks, queues, or downstream agents on the output side.

kubectl apply -f canvas.yaml · or drag on the canvas.

Alert source
alert / incident
pagerduty · webhook
Agent · runbook-enhancer v1
runbook-enhancer
2/2 pods · enriches + acts
Slack
#on-call
Ticket
jira / issue

Your first agent
is one click away.

Free tier includes 10 agents, 100k invocations per month, and every template in the public registry. Bring your own cluster when you're ready to scale.

Our Technology
01 / 02 — platform capabilities
runtime The execution engine under every agent.

Workflow runtime
with sub-millisecond queue.

Every agent runs on a typed workflow engine with a lock-free execution queue. Steps dispatch in under a millisecond, routing is deterministic, checkpoints are content-addressed. Ephemeral agents stay hot; checkpointed agents rehydrate on demand.

lock-free queue deterministic routing content-addressed checkpoints Rust + tokio
Execution queue IN-FLIGHT · ORDERED · LOCK-FREE Dispatcher ROUTE · EXECUTE · RECORD < 1 ms PER STEP Latency distribution 1 MINUTE WINDOW · ALL STEPS p50 p99 0 0.4 MS 0.9 MS 2 MS+ QUEUED 3,184 IN-FLIGHT 42 DONE / SEC 12.8 K
database One store. Four models. One query language.

Agent database,
multi-model.

ADB persists everything an agent touches — configuration as key/value, timelines as timeseries, embeddings as vectors, relationships as graphs. AQL is the single query language that joins across all four, so a skill can ask "recent events for this entity that match this intent."

key/value timeseries vector graph AQL
AQL AGENT QUERY LANGUAGE · SPANS ALL MODELS FIND events MATCH intent ACROSS kv,ts,vec,graph KEY / VAL CONFIG · STATE SESSION · PREFS TIMESERIES EVENTS · METRICS AUDIT TRAIL VECTOR EMBEDDINGS SEMANTIC SEARCH GRAPH ENTITIES RELATIONS One AQL query, composed across models NO GLUE CODE · NO SEPARATE STORES MATCH user:{id}-[:SAW]->(event) WHERE event.ts > now() - 1h AND SIMILAR(event.embedding, $intent, 0.8)
For Developers
01 / 05 — open source
BSL 2.0 Free for any use until you cross production scale.

Open source,
local-first.

Everything A3T runs is open source under BSL 2.0. Build locally with Docker, edit templates in VSCode, develop agents with Claude over MCP — standards-based, CI/CD-friendly, no cloud dependency.

B1 → B5 Scroll through the developer flow
BSL 2.0 license
Docker-native
CNCF SWF 1.0
DOCKER a3t/local:0.5.2 NODE-0 NODE-1 a3t-controller RECONCILE LOOP · OK CRDs installed 6 KINDS · WORKFLOW.A3T.IO BSL 2.0 LOCAL MACHINE NO CLOUD CALL-HOME → RUNS COMPLETELY OFFLINE LOCAL WIFI OK Your dev machine ONE DOCKER COMMAND · ZERO ACCOUNTS
1 Install

One command.
Docker only.

Local A3T ships as a single Docker image. No system packages, no package managers, no version juggling. Run it once, get a local cluster with CRDs, controller, and the agent runtime preloaded.

Requires Docker Desktop 4.25+ · macOS / Linux / WSL2.

Local install
$ docker run -p 8080:8080 a3t/local:0.5.2
2 VSCode tools

Agent YAML,
schema-aware.

The A3T extension brings schema validation, type hints, deployment, and live runtime monitoring straight into VSCode. Edit agent definitions with confidence, push to your local cluster with one click.

schema validation type hints one-click deploy live logs go-to-definition
adtech-agent.yaml — ~/a3t-local/templates
A3T Templates
templates
adtech-agent.yaml
payments-agent.yaml
canvas
rtb-prod.yaml
apiVersion: workflow.a3t.io/v1alpha1 kind: AgentTemplate metadata: name: adtech-agent version: "2.1.0" spec: durability: ephemeral budget: 80ms # ← schema validates variables: - name: strategy_pattern type: string
Deploy to local
3 Agent-assisted dev

Wire Claude to
your A3T runtime.

A3T exposes its runtime over MCP. Point Claude or any MCP client at your local cluster, generate new templates, run them against test fixtures, and ship through your existing CI/CD.

$ claude mcp add a3t-local http://localhost:8080/mcp

C
Claude
any MCP client
generate edit test
MCP
A
A3T runtime
localhost:8080
deploy inspect logs
schema check
dry-run
fixture suite
sign + push
4 Documentation

Read the technical docs.

A3T is a standards-based runtime. Everything is specified — from CRDs to AQL grammar to SDK internals. No proprietary magic, no abandoned tutorials.

Open docs
CONCEPTS
Runtime architecture
CNCF SWF execution model, ADB, Agent Canvas as CRD, reconciliation loop.
Read →
REFERENCE
AQL grammar
The full query language for agent state, memory, knowledge.
Read →
REFERENCE
CRD specifications
AgentTemplate, Agent, AgentCanvas, Resource, KnowledgeSource schemas.
Read →
GUIDE
CI/CD patterns
GitHub Actions, ArgoCD, promotion gates, air-gap packaging.
Read →
For Enterprise
01 / 04 — enterprise
k8s For platform teams.

Helm chart.
Your cluster.

A3T ships as a production-ready helm chart. Install into any Kubernetes — k3s, EKS, GKE, AKS, OpenShift. The controller speaks CRDs; your SRE team already knows how to operate it.

helm values.yaml chart a3t 0.5.2
# production install a3t: namespace: a3t-system controller: replicas: 2 resources: requests: { cpu: 500m, memory: 512Mi } registry: url: registry.yourco.com/a3t cosign: required
2 CRDs

Custom Resources
all the way down.

Every A3T object is a Kubernetes Custom Resource. Agents, templates, canvases, and every resource the agents consume — all reconciled by the A3T controller, all manageable with kubectl, ArgoCD, or your own tooling.

$ kubectl api-resources | grep a3t.io

CRD workflow.a3t.io/v1alpha1
AgentTemplate
templates.workflow.a3t.io
CNCF SWF spec
schema
Agent
agents.workflow.a3t.io
Running pod
schema
AgentCanvas
canvases.workflow.a3t.io
Composition
schema
Resource
resources.a3t.io
Typed binding
schema
KnowledgeSource
knowledge.a3t.io
Read surface
schema
MemoryBinding
memory.a3t.io
Write surface
schema
3 Admin API

Control plane,
programmable.

Every admin action in A3T is a REST/gRPC endpoint. Provision tenants, manage RBAC, install CRDs, promote resources across clusters — all from your own automation.

OpenAPI 3.1 spec · gRPC reflection · Go and TypeScript SDKs.

Admin API https://api.a3t.io/admin/v1
GET
/tenants/{id}
Get tenant metadata and quota usage
POST
/tenants/{id}/rbac/bindings
Grant role to user or service account
POST
/clusters/{id}/templates/promote
Promote a template across environments
GET
/clusters/{id}/reconcile/status
Reconciliation state for all canvases
DEL
/tenants/{id}/agents/{name}
Gracefully drain and remove an agent
4 Security

SSO, RBAC,
signed supply chain.

Enterprise deployments bring SSO via OIDC, namespace-scoped RBAC, OPA/Gatekeeper policies, and cosign-signed templates. Your security team already trusts these primitives.

  • SSO via OIDC
    Works with Okta, Auth0, Azure AD, Google Workspace, Keycloak. User claims map to A3T RBAC roles at login.
  • Namespace-scoped RBAC
    Team A authors templates in team-a-dev, ships to team-a-prod. Kubernetes handles the isolation.
  • Signed supply chain
    Templates are cosign-signed OCI blobs with SPDX SBOM. Validating admission controllers reject unsigned at apply.
  • Observability you already have
    OpenTelemetry traces, Prometheus metrics, structured logs to stdout. Datadog, Grafana, Splunk — all work.