Blueprints, runtimes, and shared threads.
Software-defined companies
Shared agents for every customer.
Talon runs durable agents with memory, tools, and policy. Build one service layer per customer instead of one assistant per seat.
apiVersion: talon.dev/v1alpha1
kind: AgentTemplate
metadata:
name: customer-support-tier-1
namespace: acme-corp
spec:
features:
- name: customer_id
type: string
required: true
model:
provider: openai
name: gpt-4o
temperature: 0.2
systemPrompt: |
You are a customer support agent on behalf of Acme Corp.
mcpServers:
- name: stripe-readonly
url: http://stripe-mcp.local/sse
execution:
memoryScope:
tenant: "{{ features.customer_id }}" Context survives across sessions and teams.
Support, success, and ops live in one system.
Templates, tools, and policies are first-class objects.
Why Talon
Most agent stacks stop too early.
Frameworks help you build one assistant. Talon helps you run a shared system of agents.
Chat UI, agent harness, orchestration layer, MCP glue, and app code drifting apart.
One control plane for auth, workers, memory, tools, policies, triggers, and sessions.
Architecture
A worker engine with a gateway, state store, and broker.
Gateway for entrypoints. Workers for execution. Database for state. Broker for streaming.
server:
host: "0.0.0.0"
port: 18789
control_plane:
database:
driver: postgres
message_broker:
driver: gcp_pubsub
manifests:
path: "./manifests" | Gateway | Auth and routing. |
| Workers | Turns, tools, and workflows. |
| Database | Templates, agents, sessions, memory. |
| Broker | Tokens and events. |
Ontology
Everything important is an object.
Agents should be inspectable, versioned, and composable.
Context, boundaries, and guardrails.
Integrations, graphs, and event bindings.