EXHIBIT 07 / AGENT SYSTEMS
Who keeps an agent’s loop turning—and decides when it stops?
The harness is the turn-by-turn program around a model call. It assembles context, invokes the model, validates tool requests, handles results, persists messages, and decides whether to continue, pause, or stop.
THE SHORT ANSWERThe model supplies outputs. The harness owns the run.
5–8 MIN · MANUAL CONTROLS · NO ACCOUNTOne loop has jobs the model cannot perform by writing text.
- 01 / ASSEMBLEBuild this turn’s input
The harness chooses which instructions, messages, memory, tool descriptions, and results the model receives.
- 02 / INTERPRETClassify the output
A response may be user-facing text, a tool request, invalid structure, or a signal that another software rule must handle.
- 03 / ORCHESTRATERun controlled work
The harness connects validation, approvals, execution, result delivery, streaming, retries, and cancellation.
- 04 / PERSISTMake the run recoverable
Messages and execution state can be stored so software—not the model’s hidden state—can resume after a disconnect.
“Harness” names a mechanism layer, not one required product. Some systems combine it with runtime, UI, memory, and scheduling; others expose a smaller loop that callers assemble themselves.
Follow the control hand-offs in a single tool turn.
This record isolates ownership. It is not a transcript from Cloudflare or any model provider, and nothing advances automatically.
HARNESS
Assemble the envelope
The program selects the system instruction, recent messages, relevant memory, and only the tools enabled for this run.
The model sees an input assembled by software.
Keyboard: focus the trace and use ← or →. Nothing advances automatically.
Operate the three gates in front of a tool.
The browser checks the gates in order and stops at the first failure. This is a deterministic rule demonstration, not a security product.
SCENARIOCandidate request: sendInvoice({ customerId: “C-104”, amount: 850 }).
DECISION / 2 OF 3 GATES ON
The action must not proceed.
Reject: executable code must not receive unchecked arguments.
A production harness also needs authentication, authorization scope, durable idempotency, failure recovery, observability, and a trustworthy tool implementation.
The source defines the layer; the museum exposes its consequences.
The harness owns turn-by-turn work
Cloudflare’s documentation lists prompt construction, model execution, tool orchestration, message persistence, streaming, recovery, and continuation among harness responsibilities.
The three-gate verdict is real
When you change a switch, this page really recomputes the first failing rule and updates the result. It does not execute the displayed tool.
Complete production safety
Three switches cannot establish correct authorization, safe retries, reliable persistence, or the behavior of any named agent product.
OFFICIAL & PRIMARY SOURCES