Agent harness
TemplatesArchitecture
Everything around an agent loop: guards, system prompt, session memory, context manager, a code sandbox, tools and an MCP server, a human approval step, and traces into an eval suite.
What is inside
- Agent ↔ model tool loop
- Tools, MCP and sandbox
- Human approval and evals
Use this template
Opens in the playground, no account needed. Sign up to save it.
Show the Inkline text
diagram architecture "Agent harness"
direction right
client app "Chat app"
zone harness "Agent harness" {
guardrail inguard "Input guard"
agent agent "Support agent"
guardrail outguard "Output guard"
prompt system "System prompt"
memory session "Session memory"
box context "Context manager"
box sandbox "Code sandbox"
}
llm model "Chat model" use=anthropic/claude
zone tools "Tools" {
tool orders "Search orders"
tool refunds "Issue refund"
mcp github "GitHub MCP"
}
person approver "Human approver"
tracer traces "Traces + cost"
eval evals "Eval suite"
app -> inguard : message
inguard -> agent
agent -> outguard
outguard ~> app : reply
system -- agent
session -- agent
context -- agent
agent <-> model : tool_use / tool_result
agent -> orders
agent -> refunds
agent -> github
agent -> sandbox : run code
refunds -> approver : needs approval
agent ~> traces
traces ~> evals