Agent Presets
A preset is a ready-made agent: a system prompt, sensible settings, and the skills it needs. Pick one with preset: and you only choose the model.
halo
agent Sorter [openai/gpt-4o-mini, preset: Classifier]
system "Labels: billing, bug, feature_request, other."
agent Reviewer [anthropic/claude-sonnet-4-5, preset: CodeReviewer]
agent Gate [anthropic/claude-haiku-4-5, preset: Judge]
pipeline Route(email: String) -> Result:
email
|> Sorter.classify() as kind
|? (context.data.kind.confidence >= 0.7)| Preset | Does | Settings | Skills | Answers |
|---|---|---|---|---|
Summarizer | Short, faithful summaries that keep names, numbers and decisions | temp: 0.2 | text | |
Classifier | Picks exactly one label from the ones you give | temp: 0 | { label, confidence, reason } | |
Extractor | Pulls the fields you ask for out of text, null when absent | temp: 0 | data-extraction | the requested fields |
Judge | Grades work against your criteria | temp: 0 | { pass, score, reasons, fixes } | |
Researcher | Searches and reads with its tools, answers with cited sources | temp: 0.3, max_steps: 20, tools: auto | web-research | text with sources |
Writer | Clear prose for the audience you name | temp: 0.6 | text | |
Editor | Tightens and corrects text, keeping meaning and voice | temp: 0.2 | text | |
Planner | Breaks a goal into ordered, checkable steps | temp: 0.2 | { goal, assumptions, steps: [{ id, task, depends_on, done_when }], risks } | |
Coder | Changes code with its tools in small, tested steps | temp: 0, max_steps: 40, tools: auto | write-tests | a summary of what it changed |
CodeReviewer | Reviews a diff for bugs and security problems | temp: 0 | code-review | { verdict, summary, findings } |
Translator | Translates into to:, keeping formatting and placeholders | temp: 0.2 | text | |
Triager | Triages bug reports: kind, severity, duplicate, what's missing | temp: 0 | bug-triage | { kind, severity, reproducible, duplicate_of, area, labels, missing, summary, reply } |
Evolver | Rewrites a Halo pipeline to do better on its criteria, inside its fence (used by halo evolve) | temp: 0.8, max_tokens: 8000 | Rationale: line and a ```halo block |
halo agents prints this list. Preset names are case-insensitive, and halo validate reports one that doesn't exist.
Changing a preset
Anything you write on the agent wins:
- Settings in the brackets replace the preset's:
[openai/gpt-4o, preset: Writer, temp: 0.9]. systemis added after the preset's prompt, so use it for what's specific to you: the labels to choose from, the criteria to judge by, your audience, your codebase's conventions.skills:adds to the preset's skills:[..., preset: CodeReviewer, skills: "house-style"](see Skills).equipsandrequireswork as on any agent.ResearcherandCoderusetools: auto, so they can call every tool they equip; give them tools and a tool policy to fence them in:
halo
policy ReadOnly : AllowEffects(read)
tool Web : WebSearch()
tool Reader : Browser(maxChars: 8000)
tool Repo : Git(".")
agent Scout [anthropic/claude-sonnet-4-5, preset: Researcher]
equips Web, Reader
requires ReadOnly
agent Reviewer [anthropic/claude-sonnet-4-5, preset: CodeReviewer, tools: auto]
equips RepoCoder is best given a computer to work on: [..., preset: Coder, sandbox: "e2b"] gives it a fresh cloud sandbox for each run, and sandbox: "docker" a container on this machine (see Remote Computers).
Presets that answer JSON tell the model the exact fields, and Halo parses JSON answers, so gates can read them directly: |? (context.data.verdict.pass == true). To make a pipeline fail when the shape is wrong, give it a schema return type.