Agents
Agents are specialized subprocesses that Fellowship spawns to handle specific roles. Unlike skills (which are prompts), agents run as independent Claude Code instances with their own context windows.
The Fellowship's Agents
Palantir
Background MonitorA thin reporter over the CLI's health sweep during fellowship execution —
runs fellowship health and reports its stalled, zombie, and
struggling classification rather than recomputing it. Reads each quest's
history for scope drift (files touched outside the described scope) and file
conflicts between parallel quests. Reports issues to Gandalf for intervention.
haiku model by
default (configurable via models.palantir). Configure via palantir.enabled and palantir.minQuests in settings.
Below the threshold, or with palantir disabled, Gandalf runs the same health
sweep itself — monitoring never depends on the extra agent.Balrog
Adversarial ValidatorAttacks the implementation before anything else reviews it. Analyzes the quest diff for failure modes, writes targeted test cases using the project's existing test framework, runs them, and delivers a severity-ranked findings report. Critical and High findings must be fixed before the quest opens its PR.
Scout
Research AgentAutonomous research agent for questions that need investigation but not code changes. Explores the codebase, reads documentation, and optionally spawns a fresh adversarial validator subagent to verify findings. Produces structured reports with confidence levels.
scout: <question> in fellowship descriptions.
No gates, no hooks, no commits. Uses the sonnet model by default
(configurable via models.scout).Validator
Findings VerifierRead-only adversarial verifier spawned by scout during its Validate phase. Re-reads every referenced file and line range, challenges assumptions, and reports each claim as CONFIRMED, CONTESTED, or UNVERIFIED.
sonnet model by default (configurable via models.validator).