Agent catalog

Agents in Beakr are not class hierarchies -- they are stateless configs. Each one is a stateless config that names a model, a system prompt, a tool allowlist, and a handoff template. The same react graph compiles around any of them.

Each agent is a (model, prompt, tool allowlist, handoff template) tuple. The graph is identical -- only the config changes.

Reference

RESEARCHER

Searches across all connected sources -- knowledge base, Slack, Gmail, Drive, Jira -- and returns cited excerpts. Designed for breadth, not depth.

ANALYST

Deep reasoning, hypothesis generation, and REPL-driven exploration. Used when a question requires multi-step logic or cross-referencing several sources.

CODER

Computation, data pipelines, and plotting. Runs code in a sandboxed environment with full package access.

REVIEWER

Verification, fact-checking, and citation auditing. Cross-checks claims against source material and flags unsupported statements.

How agents work together

The parent agent decides whether a task should stay local, be delegated to one specialist, run through a sequential pipeline, or fan out to several agents in parallel. Parallel agents are useful when the work can be split cleanly, such as one researcher gathering sources while a reviewer checks citations. Pipeline agents are useful when each stage depends on the previous stage's output.

Why stateless configs

Composable. Pipelines, war rooms, and debate ensembles can slot any config into any position.
Swappable. Swapping researcher from haiku to sonnet is a config change, not a code change.
Testable. Configs are plain data -- serialize, diff, snapshot-test.
Versioned. Every config has a hash; checkpoints record which config produced them.