Opencode Agent Creator

OpenCode Agent Creator

Build precise, functional, professional OpenCode agents and agent teams without re-researching the platform. The reference shape is one visible primary "supervisor" that plans and delegates, dispatching a roster of model-tiered subagents through the native task tool. This skill encodes the official OpenCode agent model plus field-tested orchestration, topology, and cost-routing patterns.

Scope: this skill creates OpenCode agents (the runtime actors). For repo instruction files use agent-md-creator; for portable Agent Skills use skill-creator; for Claude Code subagents use agents-claude-creator. An OpenCode subagent often loads a Skill to get its methodology — the two compose via the skill tool.

The one constraint that shapes every decision

OpenCode fixes the model per agent. The native task tool accepts only description, prompt, and subagent_type — there is no model parameter. A primary agent cannot pick a model per dispatch at runtime. Therefore:

• Cost control = routing, not runtime selection. Pre-create agents pinned to different models; the supervisor controls cost by choosing which agent to dispatch — a cheap model for trivial/bulk work, a strong model for hard reasoning, each a separate agent.
• "Create a team with the best/cheapest models" is a config-time act (writing model: into each agent), not something the supervisor does live.
• To get "cheap for easy, strong for hard" inside one domain, make two agents (e.g. x-operator and x-operator-heavy) and let the supervisor route between them.

Internalize this first — it is why the architecture looks the way it does.

The cold-context contract (the other thing that shapes every prompt)

A dispatched subagent starts cold. It sees only the dispatch prompt — not the conversation, the files the primary read, the skills the primary loaded, or the active agent mode. Any skill name, target context, constraint, or behavioral instruction must be written into the dispatch packet or it does not cross. Design each subagent prompt as if briefing a brand-new contractor who can read the repo but was in none of your meetings.

Workflow

1. Single agent, or a team — and which topology?

• Single subagent — one bounded specialty (a reviewer, a researcher). Write one .md; skip the orchestration sections.
• Team — multi-domain work, parallelism, or cost tiering. Default when the user says "agents" (plural), "team", "squad", "orchestrator", "swarm", or "sub-agents".

Pick the topology by what the work needs (full catalog + native-support status: references/multi-agent-topologies.md):

2. Don't recreate the built-ins

OpenCode ships agents — extend or route to them instead of cloning:

• Primary: build (all tools), plan (edits + bash gated to ask). (compaction, title, summary are hidden system primaries.)
• Subagent: general (full tools except todo; the docs' own answer to "run multiple units of work in parallel"), explore (fast read-only codebase search), scout (read-only external-docs/dependency research).

Create a custom agent only when a focused role, tool boundary, or pinned model beats the built-ins.

3. Choose where agents live, and scaffold

• Markdown (preferred for prompts): ~/.config/opencode/agents/<name>.md (global) or .opencode/agents/<name>.md (project). Filename = agent name (no name field in frontmatter). Frontmatter + system-prompt body.
• JSON: the agent key in opencode.json (terse model/permission overrides).
• Both merge; Markdown is auto-discovered. Keep one source of truth per agent to avoid drift.
• Scaffold fast with opencode agent create (interactive) or non-interactively with --path --description --mode --permissions --model. List with opencode agent list.
• Make the visible primary the launch default with "default_agent": "<supervisor-name>" in opencode.json (must be a primary).

Full field reference: references/agent-config-reference.md. Copy-ready skeletons: assets/.

4. Design the roster (team case)

Separate by independent decision boundary, not convenience:

• 1 supervisor — mode: primary, the only CLI-visible agent. Plans, gates scope, decomposes, dispatches, reviews evidence, synthesizes. Does not execute noisy/destructive work itself.
• N specialist subagents — mode: subagent, hidden: true. One bounded domain each, pinned to a capable model.
• M utility subagents — mode: subagent, hidden: true. Trivial/bulk/token-heavy work (summarize, parse, public lookup, boilerplate), pinned to cheap/free models.

Keep each subagent to one bounded question with one deliverable. If you cannot state its success signal in one sentence, split it. A scout-first habit (dispatch a read-only explore/scout to gather context before any write) keeps expensive agents cheap — see the OpenAgentsControl-derived discipline in references/multi-agent-topologies.md.

5. Assign models by tier (cost routing)

Model id is always provider/model-id (e.g. github-copilot/gpt-5.4, opencode/big-pickle). Leaving a subagent's model unset makes it inherit the primary that invoked it — the native "fallback" and the single dial for the whole team. The Zen catalog, free tier, inheritance-as-fallback, the per-call-model plugin workaround, and the free-model privacy caveat are in references/model-tiering.md.

6. Wire permissions and the no-cascade guard

• Supervisor — whitelist only its roster so unrelated agents vanish from its Task tool (last matching rule wins; put * first):

  permission:
    task:
      "*": deny
      "specialist-*": allow
      "utility-*": allow
  

  If it should dispatch-and-inspect but never execute, add edit: deny and bash: deny (it keeps read/grep/glob/list to review results).
• Every subagent — permission: { task: deny }. OpenCode has no recursion depth guard (issue #18100), and subagent peer-delegation is unreliable anyway (issue #7296). Keep delegation strictly hierarchical: only the supervisor dispatches. Without task: deny, a subagent can spawn subagents and create runaway sessions.

7. Write the prompts

• Supervisor body: authorization/scope gate → roster table → dispatch protocol → anti-loop + delegation discipline → routing table. The single most important rule: subagents start cold and inherit nothing but the dispatch prompt. Any skill name, behavioral mode, target context, or constraint must be written into the packet or it does not cross.
• Subagent body: role + which Skills it loads + mission + operating principles + evidence requirements + a tight output contract. Skill names are bare (no paths); the subagent loads them via the skill tool.

The body shape, dispatch protocol, anti-loop/delegation rules, and parallel-batch discipline: references/orchestrator-pattern.md. Two prompt styles work — a lean imperative body (Claude-style) or the XML-structured <role>/<task>/<workflow> body OpenAgentsControl uses; pick by team convention, both are covered. Ready-to-adapt skeletons: assets/.

8. Add tools and plugins (optional, high-leverage)

• Custom tools (~/.config/opencode/tools/*.ts): TypeScript functions the agent calls like built-ins; can shell out to Python. Best for deterministic wrappers (recursive web research, output parsers, redaction) that beat an LLM on cost and reliability.
• Plugins (plugin array or ~/.config/opencode/plugins/): npm or local; hook lifecycle events. High-value: opencode-background-agents (async read-only delegation that survives compaction), context pruning, persistent memory.

APIs, the ecosystem plugin shortlist, experimental flags, and a worked deep-research tool: references/plugins-and-tools.md.

9. Handle secrets correctly

Never hardcode keys/tokens in opencode.json — use {env:VAR} or {file:~/.secrets/x}. OpenCode does not auto-load .env, and {env:} resolves at startup before plugins, so the variable must already exist in the environment. On Windows, a small launcher that loads a .env into the process then runs opencode is the reliable path — see references/plugins-and-tools.md.

10. Validate and test

• Each agent's frontmatter is valid YAML between ---; name/filename match; model ids are real (opencode models). opencode.json parses if used.
• Restart OpenCode after edits so frontmatter is re-read.
• Smoke-test: @mention a subagent directly (behavior), and give the supervisor a task that should fan out (routing). Verify it dispatches (not executes), respects the task whitelist, and that hidden agents stay out of the @ menu. Full method: references/triggering-and-testing.md.

Anti-patterns

• Hardcoding a per-task model in the supervisor prompt — impossible natively; route to a pre-created pinned agent.
• Forgetting task: deny on subagents — invites recursive fan-out (no native depth guard).
• Building a true mesh (subagents calling each other) — unreliable on OpenCode; emulate hub-and-spoke through the supervisor.
• Assuming active skills/modes propagate to subagents — they don't (cold context); inject into the packet.
• Recreating build/plan/general/explore/scout — extend or route to the built-ins.
• Specializing portable Skills on OpenCode internals — keep model-tiering/Task-tool details in the agent files, not in reusable Skills.
• Over-tiering — don't split into cheap/strong variants unless the cost delta is real; one well-chosen model is simpler.
• Secrets in config — always {env:}/{file:}.

Resources

• references/agent-config-reference.md — complete frontmatter + opencode.json field reference: mode, model, hidden, temperature, top_p, steps, reasoningEffort, color, disable, the full permission.* key list and permission.task globs, tools (legacy), default_agent, model variants, the built-in agents, opencode agent create/list CLI, and session navigation. Load when writing or debugging agent config.
• references/multi-agent-topologies.md — the topology catalog (subagent, supervisor/hierarchical, swarm/parallel, mesh, template), each with when-to-use, OpenCode native-support status, a config sketch, and the OpenAgentsControl-derived discipline (scout-first context, dependency-batched parallelism, context-level allocation, stage gating). Load when choosing or composing a team shape.
• references/orchestrator-pattern.md — supervisor + subagent architecture on the native task tool: adaptable prompt skeletons (lean and XML-structured), dispatch protocol, wave parallelism, anti-loop guard, delegation discipline, cold-context packet rules. Load when building a team.
• references/model-tiering.md — the per-agent-fixed-model constraint, the Zen catalog and free tier, cost-routing tiers, inheritance-as-fallback, the dynamic-model-selection roadmap, and the free-model privacy caveat. Load when assigning models or controlling cost.
• references/plugins-and-tools.md — custom tools API (TypeScript, Zod args, context), plugin hooks/events, the ecosystem plugin shortlist (opencode-background-agents, workspace, skillful, supermemory), experimental flags, a worked deep-research custom tool, and the secrets/.env handling pattern. Load when extending agents with tools, plugins, or env-based secrets.
• references/triggering-and-testing.md — does the supervisor route to the right subagent, and does each agent stay in its permission boundary and honor its output contract? Triggering, behavior, and adversarial checks. Load before finalizing.
• assets/agent-template.md, assets/supervisor-template.md — fill-in skeletons for a subagent and a primary supervisor. assets/examples/ — ready-to-adapt agents: supervisor.md, code-reviewer.md (read-only), researcher.md (scout-style), summarizer.md (cheap utility). Read for concrete patterns.