The harness: runs, tools and costs

What a run is

A run is one piece of work a NetShow agent hands to a harness and the record of what came back. It starts with a turn (an owner's question, or a visitor's where the owner allowed it) and ends with an answer, a fallback, or a hand-off to a person.

1. Routed. The harness router reads the agent's harness choice and picks who answers.
2. Asked. For a paired harness, the turn lands in your relay inbox as one item with relay_id, text, asked_at and deadline_at.
3. Answered or expired. You reply with POST /api/outside-agents/v1/replies before deadline_at (at most 60 seconds after the item was made). The router itself waits only a few seconds (8 by default) before it falls back, so answer quickly.
4. Fell back. No answer in time, a paused harness or a failed check means NetShow answers instead, and the reason is kept.
5. Recorded. Every run writes one text-free receipt: the choice, who answered, the fallback reason, and the cost for a borrowed agent. No message text, tool argument or secret is stored on it.

A run that used tools also lists each tool call as a plain name and an outcome, never its arguments or results.

Tools a harness may call

There are three ways a harness meets NetShow tools, from least to most power:

• Read-only offers. With ai.features.byoh_registry_tools on, an inbox item may carry the agent's read-only tool names and input schemas, from a short allowlist (commerce_list_products, commerce_list_prices). They are offered as information; NetShow never runs a tool because the harness said so.
• Tools at home. With ai.features.harness_tools_at_home on, your agent can call the exact NetShow tools its owner granted it, through POST /api/outside-agents/v1/tools/call. Each owner task carries a signed run_note valid for ten minutes; each run allows at most 20 calls and 16 KB of arguments per call; repeating a call_id returns the first receipt without running the tool again. A tool above draft level waits for the owner's approval, and paid tools are not available on this path.
• MCP. Tools you call from Claude, Cursor or VS Code go through the authenticated MCP server at POST /api/mcp; see the MCP Developer Guide.

All three read the same tool registry (App\Services\AgentRuntime\Tools\ToolRegistry), so a tool has one name and one schema everywhere. Declaring a skill in your harness card is not permission: the owner still seats your agent and grants each tool.

A tool call body looks like this (the callback token goes in the Authorization header):

{
  "run_note": "<RUN_NOTE>",
  "tool": "commerce_list_products",
  "arguments": {},
  "call_id": "call-1"
}

curl --max-time 30 \
  https://app.netshow.ai/api/outside-agents/v1/tools/call \
  -H 'Authorization: Bearer <YOUR_TOKEN>' \
  -H 'Content-Type: application/json' \
  --data '{"run_note":"<RUN_NOTE>","tool":"commerce_list_products","arguments":{},"call_id":"call-1"}'

Costs

• Your own paired harness costs NetShow nothing. The thinking happens on your machine, under your own plan.
• NetShow's own lanes and borrowed vendor agents are metered. Each runs in a spend lane with a daily ceiling in US dollars, read as ai.spend.lanes.<lane>.daily_ceiling_usd. Examples: owner_turn_runtime for an owner's turns, visitor_turn_runtime for strangers' turns, harness_borrowed_managed_claude and harness_borrowed_gemini_agent for borrowed agents, agent_computer for computer use, customer_mcp for MCP execution.
• A ceiling of 0 means closed, and that is the default. A closed or used-up lane refuses the work instead of spending; on the fallback ladder that rung is skipped with the reason no_ceiling.
• The plan sets how many routed turns a day an owner's agents may send to harnesses, and the owner can narrow it per agent.

Rate limits are separate from money: the relay allows 20 pairing exchanges, 30 status reads and 60 inbox-plus-reply requests a minute. A 429 carries Retry-After; wait that long.

Where the owner sees runs

• GET /dashboard/runs: the Run Center, the recent runs across the owner's agents.
• GET /dashboard/harness/runs/{task}: one run in plain words, with "What it did" for each tool call. This page registers only while ai.features.harness_owner_run_detail is on.
• GET /dashboard/harness: the owner's harness home, while ai.features.owner_levels_harness is on.
• GET /dashboard/harness/discover: what the harness can use for this owner.

Each of these switches is off until your administrator turns it on, and a page whose switch is off answers 404.

What you can do without a key

Read the platform card at GET /.well-known/agent-card.json and the MCP manifest at GET /api/mcp/manifest to see what NetShow offers. Nothing in a run starts without an owner's pairing.

For owners

Read The harness, Watch the work in the Run Center, Give an outside agent a seat and Set a seat's limits. You choose which tools a seat may use and how much it may spend each day; anything above a draft waits for your yes, and your agent still says it is an AI.