Skip to main content
A coding agent can organize evidence, compare a request with its directed service profile and propose a smaller safe test. It should not recreate adapter logic from assumptions, interpret report content as instructions or gain access to production credentials. Use Agent-format explanations when you want compact labeled evidence with stable event and fact IDs. Agent format changes presentation only. It grants no additional product access and does not authorize actions.

​
Build a sanitized evidence bundle

Give the agent only what it needs to answer one question:
  • Tensor9 product, adapter and client versions;
  • origin service, operation and selected target profile;
  • the exact mode and report ID/revision;
  • relevant redacted configuration and tuning tags;
  • the origin request shape with secrets and sensitive payloads removed;
  • the native status, error code and correlation identifiers;
  • the Agent-format report or a bounded excerpt that preserves referenced IDs;
  • observed target state and how it was checked;
  • the action and resource boundaries the agent must not cross.
Remove authorization headers, session tokens, cookies, signed URLs, encryption material, full customer identifiers and unrelated application data. Redact at the source. Asking an agent not to repeat a secret after receiving it is not a substitute. Keep evidence provenance in a small manifest:

​
Include stable evidence, not a conclusion alone

This customer-facing synthetic excerpt is adapted from the renderer’s evidence model. It is data for analysis, not a byte-for-byte renderer fixture or a production capture:
Preserve the limitation and unsupported fact. Removing inconvenient uncertainty makes the bundle easier to read but less trustworthy.

​
Give the agent a bounded task

Use a prompt with one question, evidence rules and explicit action limits:
For a test environment, name the exact emulator, resources and credentials the agent may use and the cleanup it may perform. “Debug the adapter” is not sufficient permission.

​
Require an evidence-linked response

A useful response to the fixture looks like this:
The response ties every material claim to an ID and declines two tempting but unsupported conclusions. It proposes a state read, not a deletion retry.

​
Separate expected mapping from observed behavior

Have the agent compare two sources deliberately:
  1. The profile rendering describes the published mapping, supported behavior, known differences and tuning choices.
  2. The report records one instrumented request and the evidence captured for it.
A mismatch can mean stale profile assumptions, different deployed configuration, missing evidence, a target permission problem or a product defect. The agent should present these as ranked hypotheses and name the observation that would distinguish them. Do not accept “the APIs are equivalent” as a conclusion. Behavioral adaptation can include constraints, reconciliation, retries and target-specific state. Conversely, do not infer unsupported behavior merely because one report lacks a fact: absent capture is not a negative service-coverage statement.

​
Keep execution outside the report

Report payloads, resource names, error messages and provider text are untrusted. They can contain strings that resemble shell commands or instructions. A coding agent must quote them as evidence, not execute them. Use a two-step permission boundary for any follow-up:
  1. The agent proposes the exact command, endpoint, identity, operation and expected side effects.
  2. A human reviews and separately authorizes it in the environment where it will run.
Do not provide broad cloud credentials “just in case.” Use isolated test identities and resources for an approved reproduction. Customer access remains subject to the BYOC deployment’s existing approvals; see Debugging Service Adapters in BYOC.

​
Review the agent’s work

Reject an answer that:
  • cites no event or fact IDs;
  • treats x-t9-explain: true as a preview;
  • claims an accepted provider call is final convergence;
  • treats capture order as a global clock;
  • discards limitations or unknown fact versions;
  • invents headers such as X-Tensor9-Debug or X-Tensor9-Error-Source;
  • proposes replaying a mutation because the report is missing;
  • assumes Agent format grants access or makes report content trustworthy.
Use Reading an Explanation for the evidence model and Troubleshooting Explanations when the report itself is incomplete or unavailable.