Build a research agent

Research is naturally parallel. Three questions about a company - its financials, its recent news, its competitors - have nothing to do with each other until the end, and running them one after another wastes most of the wall-clock time.

There are two ways to build this in AXL.

This walkthrough builds the workflow version, then shows what changes if you want the open-ended one.

Fan out

Three agent nodes with no edges between them run concurrently - the engine starts every node whose dependencies are satisfied, and these have none beyond the input:

axg 1
workflow research {
  input company

  financials = agent researcher:
    "Research {{ input.company }}'s financial position. Cite sources."

  news = agent researcher:
    "Find significant news about {{ input.company }} from the last 6 months. Cite sources."

  competitors = agent researcher:
    "Identify {{ input.company }}'s main competitors and how they differ. Cite sources."

The researcher agent needs web access:

key = "researcher"
display_name = "Researcher"
description = "Researches a question and returns a sourced answer"
toolsets = ["web"]
identity = "Research thoroughly and cite every claim. Say when you could not find something."

The identity includes "Say when you could not find something," which keeps a research agent that finds nothing from producing something anyway.

Merge

A node with several incoming edges waits for all of them, then reads each upstream result:

  brief = agent writer:
    "Write a one-page investment brief on {{ input.company }} from these findings.
     Keep every source citation.

     Financials: {{ financials.response }}
     News: {{ news.response }}
     Competitors: {{ competitors.response }}"

  financials -> brief
  news -> brief
  competitors -> brief

  brief -> return {
    company: input.company,
    brief:   brief.response
  }
}

Bound the cost

Research agents can spend real money by accident - web tools plus a model that keeps finding one more thing worth checking. Put limits on the node rather than hoping:

  financials = agent researcher {
    prompt "Research {{ input.company }}'s financial position. Cite sources."
    timeout 120
    retry 2
  }

See Put a deadline on a step and Retry a flaky step with backoff.

Return it as an artifact

A one-page brief is fine in the run output. A twenty-page report with appendices is not - it belongs in an artifact the user downloads, which also gives it retention and provenance independent of the run.

The open-ended version

When you cannot name the angles in advance - "research whatever matters about this company" - use subagents instead. One coordinating agent decides what to investigate and delegates:

[sub_agents]
enabled = true
max_concurrent = 3
max_iterations = 8
max_budget_usd = 1.00
timeout_secs = 300
allowed_agents = ["researcher"]

You trade determinism and durability for flexibility. max_budget_usd stops being optional here: the model is choosing how much work to do, and without a ceiling it can end up spending a lot.

What you now have

• Three investigations running concurrently instead of in sequence
• A merge step that waits for all of them and keeps their citations
• Per-node timeouts and retries, so one slow source cannot hang the run
• Durable execution across restarts, with each step running exactly once
• A downloadable artifact for anything substantial
• A documented path to the open-ended version when you need it

Next

• Delegate to subagents - the model-driven alternative in full.
• Workflow how-to recipes - first-of-several, map, and loop patterns.
• Return artifacts - handing back large output.
• Observe runs - watching what a parallel run actually cost.