Outrider Agent

The Outrider Agent: a read-only migration specialist on the Federal Frontier harness, running on a local Gemma 4 26B model and grounded in Outrider assessment tool calls.

The Outrider Agent is a migration specialist on the Federal Frontier harness. It reads the estate through Outrider’s assessment tools and turns the results into briefings, wave plans, approval packets and failure diagnoses, with every fact cited to a tool call. It runs on a local Gemma 4 26B model served by llama.cpp, so no data leaves the customer’s hardware. The agent works in the Migration Monitor panel, and a person approves every move: the agent never moves anything.

Read-only identity

  • The agent authenticates as the Keycloak client outrider-agent, which holds only the migration-viewer role.
  • Outrider returns 403 to any submit or approve request from the agent.
  • The agent can only call Outrider’s read-only assessment tools through Remuda.

Tasks

All five tasks run in the Migration Monitor’s Outrider Agent panel. Every fact carries a citation to a tool call, and the tool-call evidence is viewable in the panel.

Task Output
Readiness briefing A briefing on the estate’s migration readiness.
Wave plan A plan grouping VMs into migration waves.
Approval packet A draft approval packet for one VM.
Failure diagnosis A diagnosis built from Outrider job events.
Chat Questions and answers over the estate.

Guardrails

  • Tool allowlist: the agent can call only the tools on its list.
  • Budgets: caps on the number of tool calls and on time.
  • Schema validation: output is validated against a JSON schema, with one repair attempt.
  • Grounding checks: output is rejected if it contains VM names, numbers or citations that are not present in the tool results.
  • Cross-check: results are checked against Outrider’s own can_proceed and blockers.
  • Safety-settings filter: any recommendation to change Outrider safety settings is removed.

Measured results

Measured on 2026-10-10 against a simulated vCenter estate.

Task Result Median latency
Readiness briefing 5/5 runs; blocker precision and recall 1.00; zero invented VM names, numbers or citations ~70 s
Wave plan 5/5 runs; blocker precision and recall 1.00; zero invented VM names, numbers or citations ~80 s
Approval packet 5/5 runs; blocker precision and recall 1.00; zero invented VM names, numbers or citations ~20 s
Failure diagnosis 30/30 correct on recorded Outrider job runs —
Chat 9/10 ~10 s

On a second, 3-VM estate, the agent scored 5/5 on every task.

The evaluation conditions are listed under Current scope.