Skip to content

marola-agents

The inventory of every agent marola uses or plans to use: the coding agents that write and review the code, the Claude Projects and their routines, the agents inside the product, the ones CI calls, and the ones still on paper. For each one it says where the agent lives, what triggers it, which model runs it, what it may touch and what it may not.

Status: a submodule of the umbrella since marola-dev/marola#793, published at https://docs.marola.dev/5-Repos/marola-agents/. The design MIP comes later. Inventory read on 2026-10-11 from every repo's main.

Nothing here is loaded by Claude Code or any other tool. Each agent keeps living in the repo that runs it, and this repo describes it. Skills are catalogued separately, in marola-skills.

What counts as an agent

A model that acts on its own between a trigger and a result: it reads, decides and writes a report, a comment, a commit or an answer. A script that only calls an API (the arXiv digest, the wiring generator) is not one, and neither is a bot with no model in it (Scala Steward, the pointer-sync PR).

Coding agents (Claude Code subagents)

Agent Repo Model Tools Job Writes?
mip-reviewer marola-devkit plugin fable Read, Grep, Glob, Bash Reviews one PR of a MIP task stack against its .tasks.md row; correctness gaps only No, it reports
mip-claims-auditor marola-devkit plugin haiku Read, Grep, Glob, Bash, WebFetch Finds unsourced external claims in a MIP's §4 No, it reports
jar-verifier marola-app sonnet Bash Checks a Kyo API against the pinned jar with javap No, it reports
news-fact-check marola-site opus Read, Grep, Glob, Bash, WebFetch A news post's first review: every claim against its link, pt-BR and en parity No, it reports
news-copy-review marola-site opus Read, Grep, Glob A news post's second review: language, voice, clarity No, it reports
revisor-proposta h0ffmann/ww3-gpu opus Read, Grep, Glob, Bash Scientific review of the UFRJ/DEL project proposal (ABNT, register, citations) No, it reports
security-reviewer h0ffmann/ww3-gpu opus Read, Grep, Glob, Bash Security review of a branch's diff: secrets, hooks, workflows, prompts No, it reports

Every one of them is read-only by design: it reports, a person or the calling session fixes.

Claude Projects and their routines

Project Settings copy Agents
Marola Agent the umbrella's CLAUDE_PROJECT.config.md A coordinator session in the project chat and one thread session per task
Wave Forecaster h0ffmann/ww3-gpu .claude/project/ (ADR-0005) The same shape, for the WW3-on-GPU book and lab
Routine Project Schedule Job
A2A inbox (para:marola) Marola Agent every 3 h Answers [A2A] issues addressed to marola in marola-dev/marola and h0ffmann/ww3-gpu
Preparar post LinkedIn Marola Agent Mon–Wed 9:52 (São Paulo) Reminds the day before a post and suggests a topic
Lembrete LinkedIn Marola Agent Tue–Thu 9:52 (São Paulo) Reminds at posting time
A2A (para:wave-forecaster) Wave Forecaster every 3 h The other side of the same channel
Daily deps update Wave Forecaster daily 9:52 (São Paulo) Opens or drives one dependency-update PR in ww3-gpu
Weekly settings snapshot Wave Forecaster Fridays 2:30 (São Paulo) Refreshes .claude/project/ from the live settings

Agents in the product (marola-app)

Agent Where Model Job Limit
Summarizer core/…/llm/, recommendation_prompt.json a local Ollama model (llama3.2 by default, marola-sea) Writes the short summary of a beach's best hour Never changes a score or drops an alert
Reviewer core/…/llm/Reviewer.scala, review_prompt.json the same Checks the summary against the numbers Can only reject or correct the text
Ask / chat cli/…/agent/ChatServer.scala, the site's chat widget the same Answers ocean questions from the sourced corpus (RAG), with the safety footer Answers only from marola-corpus
MCP tool server cli/…/agent/SwimConditionsMcpServer.scala none of its own Exposes the pipeline to an outside agent Read-only tools
Vision core/…/vision/, local/…/vision/ a local Ollama vision model Describes a photo a user sends ("is this a jellyfish?", visibility) Answers that user only

Agents in CI

Agent Repo Trigger Job
Gemini review every marola-dev repo, via marola-devkit's gemini-review.yml a person requests the gemini team on a PR One review per request, then a fix commit only where the fix still applies, never under .github/
DSPy prompt compile marola-ml compile-prompt.yml dispatch by a person Compiles the summarizer and reviewer prompts against a model served in the job, then opens a PR to marola-app

Planned

Agent For Where it is written down State
Post writer Drafts a news post in pt-BR and en, ready for the two news reviewers marola-site's news-post skill does the writing today Idea
Site UI agents Discipline in UI work on marola.dev: design review, worst-case data, motion review, screenshots before and after marola-site's site-frontend skill orders the frontend skills today Idea
Hazard detection agent Detects hazards and compares them with public alerts umbrella FUTURE-WORK.md §9.2 Candidate, gated on a human-confirmation design
Escalation, digest and answer actors marola as actors on Pekko, each also an MCP tool umbrella AGENT-FRAMEWORKS-SURVEY.md §2–§3 Candidate
"Ask marola about marola" Q&A over the project's own docs (ask_project_question) umbrella AGENT-STACK-SURVEY.md §5 Candidate
Hosted agent over HTTP/SSE MCP The MCP server reachable from outside umbrella ARCHITECTURE.md §5c Candidate, Phase 2
Evaluation harness Per-agent scoring and cross-agent traces over the benchmark and the reviewer umbrella FUTURE-WORK.md §4.1 Candidate

A planned agent moves to its section above when the PR that adds it merges.

Rules every agent here follows

  • Read-only unless its job is to write, and then only what its row says.
  • No paid model or resource without a person's go-ahead. The product's agents run on a free local Ollama model; CI's run on free tiers or inside the job.
  • No model id, secret or session link in a file pushed to a repo. A row names the model as its definition does, never a dated id.
  • The org invariants apply as in every marola repo: the umbrella's AGENTS.md.

Agent docs already in marola

What marola has already written about agents, in the umbrella unless a repo is named.

Doc What it covers
Chat and MCP Running the chat server behind the site's widget, and the MCP server for an outside agent
Architecture Where the summarizer, the reviewer and the MCP server sit in the pipeline
Agent skills Which Claude Code skills marola uses, which repo has each, when to reach for them
Dev flow Issue to MIP to stacked PRs: the flow the coding agents follow
Gemini Code Assist The free, on-request Gemini review on PRs
Agent stack survey Three agent toolkits against the MIPs, and the design of a Q&A agent about marola itself
Multi-agent frameworks survey Multi-agent frameworks on the JVM and where Apache Pekko would fit
Awesome agentic engineering marola's own curated list of agent tools and practices (MIP-0043)
Future work The planned agents in the table above
MIP-0011 Hooks, permissions, subagents and skills in Claude Code
MIP-0017 Agentic tooling ideas taken from other repos
MIP-0040 Measuring the reviewer as a judge, then making it a small jury
MIP-0076 How an agent finds things across repos: the wiring table and the graph
Chat widget (marola-site) The site's front end for the ask agent
Prompt compile (marola-ml) The DSPy compile of the summarizer and reviewer prompts

Awesome lists

Outside lists worth reading before adding an agent here. Checked to exist and not be archived on 2026-10-11; star counts change, so none are quoted.

Keeping it current

A PR that adds, removes or changes an agent in any repo updates its row here. A row says where the agent is defined, so git grep or a look at that file checks it.

Built from marola-dev/marola-agents@9cf178a.