Agents Reference

Agents Reference

This page documents every agent in the BizarHarness pantheon. Each agent is a single Markdown file in config/agents/, with YAML frontmatter (name, model) and a Markdown body that defines its role, routing rules, and tool surface.

The pantheon at a glance

Agent	Rune	Model	Tier	Cost	Role
Odin	ᛟ	MiniMax M3	2 (mid)	$0.30/$1.20 per M in/out	Primary router — never executes, only delegates
Frigg	ᚠ	DeepSeek V4 Flash	0 (free)	Free	Read-only Q&A — answers with file references, never modifies
Vör	ᛢ	DeepSeek V4 Flash	0 (free)	Free	Clarifier — asks project-specific questions if ambiguous
Quick	ᛟ	DeepSeek V4 Flash	0 (free)	Free	Single-shot fast path — no decomposition
Mimir	ᛗ	DeepSeek V4 Flash	0 (free)	Free	Codebase research — Semble-first exploration
Heimdall	ᚹ	DeepSeek V4 Flash	0 (free)	Free	File ops, mechanical edits, quick changes
Hermod	ᚱ	MiniMax M2.7	1 (low)	$0.30/$1.20 per M in/out	Git and GitHub operations
Thor	ᚦ	MiniMax M2.7	1 (low)	$0.30/$1.20 per M in/out	Moderate implementation, debugging, refactoring
Baldr	ᛒ	MiniMax M2.7	1 (low)	$0.30/$1.20 per M in/out	Design systems, visual audits, usability plans
Tyr	ᛏ	MiniMax M3	2 (mid)	$0.30/$1.20 per M in/out	Complex implementation, architecture, deep debugging
Vidarr	ᛉ	GPT-5.5	4 (last)	OpenAI subscription	Last-resort fallback when all else fails
Forseti	ᚨ	MiniMax M3	3 (auditor)	$0.30/$1.20 per M in/out	Adversarial plan review, edit-deny

Odin (orchestrator)

• Rune: ᛟ
• Model: minimax/MiniMax-M3
• Cost: $0.30/M input, $1.20/M output
• Role: Primary router. The default agent. Decomposes every request into parallel work streams and dispatches to subagents. Never executes work itself.
• Tools: task, todowrite, Hindsight MCP tools, Semble MCP tools. No bash, glob, grep, edit, write, or question.
• When to use: Any non-trivial request that isn't a single-shot question. The default for all free-form prompts.
• Example: @odin implement the /api/export endpoint with tests
• Routing rule: Always parallel. Fires 2+ task calls in a single message. Gates Tier 4/5 work via Forseti.

Frigg (ask)

• Rune: ᚠ
• Model: opencode/deepseek-v4-flash-free
• Cost: Free
• Role: Read-only Q&A. Answers questions about the project with file references. Never modifies any files.
• Tools: read, glob, grep, Hindsight MCP tools, Semble MCP tools. No edit, write, or bash.
• When to use: "What does this project do?", "Where is X handled?", "How does Y work?"
• Example: @frigg what does this project do
• Routing rule: Direct invocation. No decomposition. Single-shot.

Vör (clarify)

• Rune: ᛢ
• Model: opencode/deepseek-v4-flash-free
• Cost: Free
• Role: Clarifies ambiguous or incomplete requests. Uses a research-first protocol — reads .bizar/PROJECT.md and the project's Hindsight bank before asking any questions. Asks only when research has been exhausted, and questions must reference actual project files, frameworks, or patterns.
• Tools: read, glob, grep, Hindsight MCP tools, Semble MCP tools, question. No edit, write, or bash.
• When to use: Open-ended requests that lack a clear target. Routes for you once the request is well-defined.
• Example: @vör build me a thing (Vör will ask what the thing is)
• Routing rule: Odin cannot ask questions. Ambiguous requests are routed here.

Quick (quick)

• Rune: ᛟ
• Model: opencode/deepseek-v4-flash-free
• Cost: Free
• Role: Single-shot fast path. Skips decomposition and delegation. Good for trivial asks that don't need parallel work.
• Tools: Full read/write tool surface, including bash, edit, write.
• When to use: One-line changes, single-file edits, "what's the answer to this specific question." Faster than Odin for tasks you know don't need decomposition.
• Example: @quick delete the unused import in src/main.ts
• Routing rule: Direct invocation. Does not dispatch to subagents.

Mimir (research)

• Rune: ᛗ
• Model: opencode/deepseek-v4-flash-free
• Cost: Free
• Role: Deep codebase research. Uses Semble as its primary search tool and falls back to grep only when Semble is unavailable. Produces structured findings with file:line references.
• Tools: read, glob, grep, Semble MCP tools, Hindsight MCP tools. No edit or write.
• When to use: "Research how X is implemented", "find every place that touches Y", "compare the auth flow across versions."
• Example: @mimir research how authentication flows through the request lifecycle
• Routing rule: Direct invocation or dispatched by Odin for the research leg of a larger task.

Heimdall (simple)

• Rune: ᚹ
• Model: opencode/deepseek-v4-flash-free
• Cost: Free
• Role: File operations, mechanical edits, quick changes. The "do the boring thing fast" agent.
• Tools: Full read/write tool surface.
• When to use: Renames, formatting fixes, simple CRUD, generating boilerplate, single-file edits.
• Example: @heimdall rename getUserById to findUserById across the repo
• Routing rule: Direct invocation or dispatched by Odin for mechanical work.

Hermod (git)

• Rune: ᚱ
• Model: minimax/MiniMax-M2.7
• Cost: $0.30/M input, $1.20/M output
• Role: Git and GitHub operations. Commits, pushes, merges, rebase, branch management, pull requests, GitHub issue and release management.
• Tools: bash, read, glob, grep, gh CLI. No edit or write (modifies the working tree via git, not via direct file writes).
• When to use: Any git operation. Any GitHub API call (issues, PRs, releases, labels).
• Example: @hermod commit the staged changes with a conventional commit message
• Routing rule: Direct invocation or dispatched by Odin. PR review uses /pr-review mode with Mimir (research) and Forseti (audit).

Thor (medium)

• Rune: ᚦ
• Model: minimax/MiniMax-M2.7
• Cost: $0.30/M input, $1.20/M output
• Role: Moderate-complexity implementation. Features that span a few files, debugging tasks, code review, refactoring of contained scopes.
• Tools: Full read/write tool surface plus all MCP tools.
• When to use: "Add a /healthz endpoint", "fix the bug in the rate limiter", "extract the validation logic into a separate module."
• Example: @thor add a /healthz endpoint that returns 200 OK
• Routing rule: Direct invocation or dispatched by Odin. Always run in parallel with Tyr on Tier 4 work.

Baldr (design)

• Rune: ᛒ
• Model: minimax/MiniMax-M2.7
• Cost: $0.30/M input, $1.20/M output
• Role: Design system creation, DESIGN.md authoring, visual audits, usability planning. Plans only — does not implement. A Baldr plan is the design contract; Thor or Tyr executes it later.
• Tools: read, glob, grep, write (to DESIGN.md and design files), MCP tools. Limited bash.
• When to use: "Build out a design system for the marketing site", "audit the visual consistency of the dashboard", "plan the color tokens for dark mode."
• Example: @baldr plan a DESIGN.md for the marketing site
• Routing rule: Direct invocation. The plan is handed off to Thor or Tyr for execution.

Tyr (complex)

• Rune: ᛏ
• Model: minimax/MiniMax-M3
• Cost: $0.30/M input, $1.20/M output
• Role: Highest-complexity implementation. Architecture decisions, cross-cutting refactors, deep debugging, multi-step engineering. Always run in parallel with Thor on Tier 4 work.
• Tools: Full read/write tool surface plus all MCP tools.
• When to use: "Refactor the auth system to use a session token", "design the migration plan for the schema change", "trace the source of the memory leak."
• Example: @tyr plan the migration from REST to tRPC
• Routing rule: Direct invocation or dispatched by Odin for complex legs. Always gated by Forseti before execution.

Vidarr (last)

• Rune: ᛉ
• Model: openai/gpt-5.5
• Cost: OpenAI ChatGPT subscription
• Role: Last-resort fallback. Invoked when Tyr stalls, debugging is stuck, or the problem is genuinely novel and the opencode model tiers have all failed. Default disabled — must be explicitly enabled in opencode.json.
• Tools: Full read/write tool surface.
• When to use: When nothing else has worked. Postmortem of failed attempts. Novel problem domains.
• Example: @vidarr figure out why the binary segfaults under load
• Routing rule: Direct invocation or dispatched by Odin. Always gated by Forseti before execution.

Forseti (audit)

• Rune: ᚨ
• Model: minimax/MiniMax-M3
• Cost: $0.30/M input, $1.20/M output
• Role: Adversarial plan reviewer. Audits completeness, correctness, consistency, feasibility, and security of plans from Tyr and Vidarr. Edit permission: denied — runs in audit-only mode.
• Tools: read, glob, grep, MCP tools. No edit, write, or bash.
• When to use: Never invoked directly. Auto-runs before any Tier 4 or Tier 5 implementation. Also runs bizarharness audit for a security review of agent config.
• Example: Not directly invokable by users; auto-triggered by Odin.
• Routing rule: Auto-dispatched by Odin for plans involving Tyr or Vidarr. Returns approve, request-changes, or reject.

Permissions summary

Agent	Read	Write	Bash	Edit	Question	Task (dispatch)
Odin	yes	no	no	no	no	yes
Frigg	yes	no	no	no	no	no
Vör	yes	no	no	no	yes	no
Quick	yes	yes	yes	yes	no	no
Mimir	yes	no	no	no	no	no
Heimdall	yes	yes	yes	yes	no	no
Hermod	yes	no	yes	no	no	no
Thor	yes	yes	yes	yes	no	no
Baldr	yes	yes	limited	yes	no	no
Tyr	yes	yes	yes	yes	no	no
Vidarr	yes	yes	yes	yes	no	no
Forseti	yes	no	no	no	no	no

Odin and Forseti are the only agents with the task tool (dispatch). All other agents are leaf nodes — they execute the work they're given and return. Vör is the only agent with the question tool — it's the system's only way to ask the user.

Next steps

Next: Model Routing — the 5-tier model architecture, per-agent configuration, and how to override models per task.