@agentcompose/analysis-agent

A configurable, spec-compliant analysis worker for AgentCompose.

It scores options against weighted criteria over supplied evidence and returns a ranked, decision-ready verdict with a recommendation — behind the capability analyze. It is a general evaluator: competitive analysis and feasibility are criteria presets, not separate agents. You can score vendors, designs, PRs, or product ideas with the same machine.

npm install @agentcompose/analysis-agent

It consumes evidence rather than gathering it — pair it downstream of the research worker for fresh data:

graph LR
    R[research<br/>gather] --> A[analyze<br/>score + rank + recommend] --> P[HITL pick]

Loading

Use

import { makeAnalysisAgent } from "@agentcompose/analysis-agent";
import { inProcess } from "@agentcompose/sdk";

const analysis = makeAnalysisAgent({
  defaults: { baseUrl: "http://localhost:4000/v1", model: "gpt-4o-mini" },
});

const client = inProcess(analysis);
await client.configure({ preset: "feasibility" });
const task = await client.submit([
  { kind: "text", text: "Evaluate the feasibility of building each as a solo dev in 4 weeks." },
  { kind: "json", json: { options: ["Habit tracker", "Multiplayer game", "Notes PWA"] } },
]);

for await (const ev of client.events(task.id)) {
  if (ev.type === "result" || ev.type === "error") break;
}
const final = await client.get(task.id);
// → [ { kind:"text", markdown decision brief }, { kind:"json", Analysis } ]

What it returns

A structured Analysis (the json part) plus a Markdown decision brief (the text part) and an analysis.md artifact:

interface Analysis {
  subject: string;
  criteria: Criterion[];
  scale: number;                // scores run 0..scale
  options: OptionScore[];       // { name, scores, notes, weightedScore }
  ranking: string[];            // option names, best first
  recommendation: { choice; confidence: "low"|"medium"|"high"; rationale };
  risks: string[];
}

Scoring is deterministic and reproducible. The model only assigns raw per-criterion magnitudes (grounded in the evidence); the worker computes the weighted, normalized total in code — honoring weights and inverted ("lower is better") criteria. The model is never trusted for arithmetic or for picking the winner.

Configuration

Key	Default	Description
preset	competitive	competitive, feasibility, or custom (then supply criteria).
criteria	—	Explicit weighted criteria; override/extend the preset by name. { name, weight?, description?, invert? }.
scale	10	Score scale upper bound (0..scale).
maxOptions	6	Cap on options when they're extracted from the evidence.
options	—	Explicit options to evaluate; omitted → extracted from the evidence.
evidence	—	Inline evidence (in addition to evidence passed in the goal).
provider	—	BYO-model: { baseUrl, model, apiKey } (OpenAI Chat Completions compatible).

Criteria & inversion

A criterion with invert: true (e.g. cost, technical effort, risk) is one where a high score is bad — the worker subtracts it from the scale before weighting, so a costly or risky option ranks lower. Presets ship sensible weights; override any by name:

await client.configure({
  preset: "competitive",
  criteria: [{ name: "demand", weight: 3 }, { name: "novelty", weight: 1 }],
});

Verticalize by configuration, not specialization

"Competitive Analyst" and "Feasibility scorer" are not separate agents — they are the same analyze capability with different criteria presets. New evaluation verticals are config, not code, which keeps the worker general and composable.

License

Apache-2.0