Skip to content

FDE 2: Environment

The environment is where the scope doc becomes operational: context (what sessions see), tools (what agents may attempt), policy (what needs approval). Build narrow, verify strictly, open up only with evidence.

Translate each scope-doc line into environment shape:

Scope doc line Becomes
In-scope systems and data Context boundaries, exactly these, no more
Out-of-scope list Explicit exclusions, checked during verification
Read and change needs Initial tool set (reads first; drafts before executes)
Risky actions Approval gates (full design next stage: Approvals)
Approver roles Who the gates route to
Connector gaps Blockers or deferrals. The environment only promises what is actually connected.

Build narrow, then verify strictly. Open up only with evidence.

  • Context covers the workload’s systems and nothing adjacent “just in case”
  • Data boundaries concrete (queues, groups, windows, not “production”)
  • Tool set minimal: reads needed now, drafts for consequential outputs, executes only where the workload requires
  • No bonus tools, no speculative scope. Pilot evidence can widen later.
  • Exclusions from the scope doc preserved as exclusions in the environment
Default Agent configuration screen in AI Agents

Verify before users arrive: the probe sequence

Section titled “Verify before users arrive: the probe sequence”

Run these as test sessions yourself, in order. Each must pass before the next:

  1. Happy-path read: the core lookup works against real data in scope.
  2. Boundary probe: out-of-scope systems and data are unreachable (attempt and confirm denial or absence, do not just assume).
  3. Draft quality: draft outputs are reviewable by the intended approver role.
  4. Gate check: a risky action fires the expected approval with what, on-what, why intact.
  5. Denial path: deny it; the session degrades gracefully instead of wedging or retry-spamming.
  6. Wrong-org check: records land under the right org scope (org_id correct).

The sequence runs: scope doc, environment spec, build narrow, then probes for reads, boundaries, drafts, gates, and denials with org scope. A failed probe loops back to the build; a full pass moves to capabilities.

Environment: <name>
Tenant: <hostname>, org scope: <org_id context>
Context: <systems and data boundaries>
Exclusions: <explicit>
Tools: <tool, kind, and rationale table>
Policy refs: <link to approval matrix>
Connector deps: <system and state, connected, requested, or deferred>
Probes: <date and results of the six probes>
  • Spec written and reviewed by a second person before pilot users enter
  • Connector deps honest: nothing marked connected that is not verified in-environment. Bindings need org, package, and connection coordinates in install settings.object_mappings (SDK concepts); registered adapters are HubSpot and ConnectWise PSA only

Exact environment-creation mechanics (console/API calls, schema): not documented in source yet. Confirm in the product repo and link from here. The operator-facing reference lives under Environments and Agent config.

Next: Skills, tools and MCP makes the environment capable.