🧙‍♂️ The agent runner

The agent runner
we're dogfooding.

Merlin is CorvidLabs's in-house AI agent runner. Multi-provider, tier-aware, plugin-shaped. Built on spec-sync contracts and the fledge plugin protocol.

Docs → Repo private during dogfooding.

Configure once, run anywhere

A real fledge.toml with Merlin enabled. Three providers, different tiers, single config file.

# fledge.toml
[merlin]
verify_before_complete = true
max_retries = 3
persona = "AGENT.md"

[merlin.providers.claude]
tier = "code"
type = "anthropic"
api_key_env = "ANTHROPIC_API_KEY"
model = "claude-sonnet-4-6"

[merlin.providers.ollama]
tier = "code"
type = "openai"
api_key_env = "OLLAMA_API_KEY"
model = "qwen3-coder:480b"
base_url = "https://ollama.com/api/v1"

What makes it different

Multi-provider

Anthropic (Claude), Ollama (local + cloud), OpenAI-compatible. Switch via fledge ai use or per-task --provider flag.

Tier-aware

Three tiers: read (cheap context builders), tool (default day-to-day), code (high-capability for writing). Different models per tier; Merlin picks based on task requirements.

Plugin tool surface

Tools are fledge plugins. Want Merlin to know how to do X? Write a fledge plugin, declare in [merlin.tools]. No fork-and-edit.

Spec-driven

Specs in specs/*.spec.md are the contract. Agents validate against them before completing a task. Drift is a CI failure, not a code-review nag.

AlgoChat-native

Spawn Merlin to listen on AlgoChat channels, @merlin mentions in corvid-chat route to it via the merlin-bridge.

Sandboxed agents (planned)

Sub-agent plugin is the next big move. Main agent stays lean, delegates heavy work to fresh-context sub-agents that return summaries. Closes the context-retention gap.

The dependency chain

Merlin sits at the start of the CorvidLabs 2026 strategy chain. The single load-bearing move is the sub-agent plugin. These are the first five steps; see /about/ for steps 6–7 (corvid-chat launch + the amplifier).