The lifecycle behind every deployment

Agent code becomes a production deployment — on purpose.

Every agent follows the same six steps on the way to production — build, package, attach, run, observe, promote. sypho runs that path as a system: versioned, isolated, and fully traced.

Build
→
Package
→
Attach
→
Run
→
Observe
→
Promote

The agent lifecycle, made operational.

Agent teams need a repeatable path from development to production, not a one-off script. sypho turns that path into a runtime workflow: package the agent, attach its tools, skills, structured data, and budget, choose where it runs, then execute and inspect every step.

Agent code
↓
Deployment manifest
↓
Tools / Skills / Structured data / Budget
↓
Runner sandbox
↓
Run timeline
↓
Debug / Deploy next version
01

The agent arrives as-is.

Start with what you already have: code, prompts, tools, workflows, product logic. sypho doesn't force a rewrite.

agent/
├─ prompts
├─ tools
├─ workflows
└─ entrypoint
02

The agent becomes a deployment.

A versioned deployment describes the runtime contract — entrypoint, tools, skills, structured data, secrets, target environment.

deployment manifest
agent: support-agent
version: 0.4.2
entrypoint: support:agent
deployment: production
03

Capabilities are attached, not assumed.

Production agents need controlled access to what's around them — not unrestricted access to everything. Skills, structured data, tool definitions, secrets, and budget are scoped individually.

Deployment
├─ tools
├─ skills
├─ structured data
├─ secrets
└─ budget
04

The runner target is explicit.

The workload decides the environment: local, private cloud, customer cloud, on-prem, hybrid, or managed execution.

Runner target
├─ local
├─ private cloud
├─ customer cloud
├─ on-prem
└─ managed
05

The runner takes it from here.

sypho resolves the deployment, selects a runner, and starts the runtime. Capabilities attach automatically; events and results stream back to the control plane.

request.received
deployment.resolved
runner.selected
agent.started
result.returned
06

Every step becomes a trace.

Each run becomes an inspectable timeline — skill activations, structured-data access, tool calls, logs, errors, retries, final results.

run.started
skill.activated
structured_data.query
tool.call
run.completed
07

Traces turn into the next version.

A trace shows exactly what happened — and why it failed, if it did. Compare versions, validate changes, and promote whichever one is safer.

trace review
↓
fix and rebuild
↓
version compare
↓
promotion

Ship agents through a production lifecycle.

Request access and help shape the runtime workflow for agentic software.