Platform
How does [product] work?
What happens when the system runs: input, reasoning, a human checkpoint, output, and a write to your system of record, with the de-risking detail at each step.
What happens, step by step?
- 01
Input
[A real signal or artifact enters: an inbound, a brief, a trigger. The agent reads the relevant kernel slice before doing anything.]
What can go wrong here: [If the input is ambiguous, the agent asks rather than guesses.]
[DOM-rendered frame for the input step: real text, no screenshot.]
- 02
Reasoning
[The agent runs its named checks, grounded in kernel evidence, and produces a confidence signal. This is where the ReasoningTrace is built, checks, evidence, confidence flag.]
What can go wrong here: [Below the confidence threshold, it flags for a human instead of proceeding.]
The reasoning trace: the deepest fixture on the site. - 01 · Input received[the signal/artifact that started the run]
- 02 · Checks performed[the named checks, grounded in the kernel]
- 03 · Evidence cited[kernel fields + external evidence]
- 04 · Confidence / flag[the confidence signal; flagged if low]
- 03
Human checkpoint
[A person, named by role, approves, edits, or rejects. What they see and what happens on rejection is explicit, not implied.]
What can go wrong here: [On rejection, nothing ships and the run is logged as declined.]
The approval surface
[What a human sees, approves, edits, or rejects, named by role. On rejection, nothing ships.]
- 04
Output
[The approved artifact is produced at full fidelity (the message, the draft, the digest) DOM-rendered so the claim is readable, not a screenshot.]
What can go wrong here: [Every load-bearing number is defensible or it does not appear.]
[DOM-rendered frame for the output step: real text, no screenshot.]
- 05
Write
[The output is written to the system of record (CRM, CMS, warehouse) with the full audit trail attached: what ran, what was cited, who approved.]
What can go wrong here: [A failed write is retried and surfaced, never silently dropped.]
[DOM-rendered frame for the write step: real text, no screenshot.]
Is [product] autonomous, and what does a human approve?
Where each agent family sits on copilot → assisted → supervised → autonomous, and what a person approves at each level. Levels differ per family and per customer configuration; this is the default. We never claim more autonomy than the product ships.
| Agent family | Autonomy level | What a human approves | What is logged |
|---|---|---|---|
| AI SDR | [supervised: TBD §12.3] | [every send before it goes to a prospect] | [input, checks, evidence, approver, output] |
| Content agents | [assisted: TBD §12.3] | [every claim + the final draft before publish] | [sources, verification result, editor, publish event] |
| Creative agents | [assisted: TBD §12.3] | [the selected variant before it ships] | [constraints applied, variant chosen, approver] |
| Research agents | [supervised: TBD §12.3] | [which findings surface to a decision-maker] | [sources, confidence, reviewer, digest version] |
How do we know it did what it says?
- 01 · Input received[the signal/artifact that started the run]
- 02 · Checks performed[the named checks, grounded in the kernel]
- 03 · Evidence cited[kernel fields + external evidence]
- 04 · Confidence / flag[the confidence signal; flagged if low]
- 05 · Human checkpoint[who approved, by role]
- 06 · Output written to[system of record · timestamp]
- [Retention: how long each run's log is kept.]
- [Export: how the log leaves the system.]
- [Access: who can query it.]
Frequently asked questions
Can we change what the agent is allowed to do?
What happens when confidence is low?
Who approves before something ships?
What is written to the audit trail?
See the machine on your own GTM
[What happens next, one line under each path: the mechanism you just read, run on your data.]