Build an agent

The easiest way to build a production-quality agent.

Assemble a lean main prompt plus proven skills from the registry — every part measured, swappable, and safety-gated.

01
Write skills in markdown
Each behavior is a SKILL.md — name, triggers, instructions. Fork from the registry or write your own.
02
Wire it in two lines
decimalai.init() captures traces and the manifest. Works with LangChain, OpenAI Agents, ADK, or plain Python.
03
Route, don’t stuff
smart_route(task) hands the agent a ranked shortlist — not your whole folder in the system prompt.
04
Eval the bundle
Run your cases with and without the skills, and keep what lifts.
05
Ship it governed
Publish to your workspace — safety-gated, versioned, shared with your team.
support-agent/
├─ skills/
│  ├─ support-triage.md · +28%
│  ├─ error-triage-protocol.md · +28%
│  └─ pii-redactor.md · +31%
├─ eval/cases.yaml · 21 cases
└─ main.py · the harness
# main.py
import decimalai
decimalai.init(langchain=True)
skills = decimalai.smart_route(task)
Bundle eval · support-agent
Pass rate without skills43%
Pass rate with bundle76%
Lift+33 pts
illustrative — run it on your own cases with the open eval runner
● Full worked example: build a support agent from skills →The general model: assemble an agent from skills →Two-minute demo →
Playbooks

Pick your agent. Start from its playbook.

Each playbook is a starter kit of proven skills, a pre-tuned router config, and a tutorial — fork the kit, or build it step by step.

Coding agents
3 benchmarked · avg +25%
Ship consistent, reviewable output from coding agents.
⌗ commit-conventions+35%
⌗ json-strict-output+22%
⌗ changelog-writer+19%
Fork the kitStart from the general guide →
per-type tutorial coming · router config included · avg lift is the mean of the skills shown
Data & SQL agents
3 benchmarked · avg +44%
Queries that return what they promise — validated, typed, safe.
⌗ sql-result-contract+66%
⌗ schema-linter+26%
⌗ postgres-migration-guard+40%
Fork the kitStart from the general guide →
per-type tutorial coming · router config included · avg lift is the mean of the skills shown
Support agents
3 benchmarked · avg +29%
Triage, resolve, and escalate without leaking a thing.
⌗ support-triage+28%
⌗ error-triage-protocol+28%
⌗ pii-redactor+31%
Build it — full tutorialFork the kit →
tutorial live in docs · router config included · avg lift is the mean of the skills shown
Security agents
2 benchmarked · avg +32%
Guard the agent itself — inputs, outputs, and secrets.
⌗ prompt-injection-shield+33%
⌗ pii-redactor+31%
Fork the kitStart from the general guide →
per-type tutorial coming · router config included · avg lift is the mean of the skills shown
Research & web agents
2 benchmarked · avg +26%
Structured extraction that survives messy pages.
⌗ web-scraper-toolkit+22%
⌗ offload-to-code+30%
Fork the kitStart from the general guide →
per-type tutorial coming · router config included · avg lift is the mean of the skills shown
Ops & infra agents
2 benchmarked · avg +35%
Changes to production that never skip the guardrails.
⌗ postgres-migration-guard+40%
⌗ offload-to-code+30%
Fork the kitStart from the general guide →
per-type tutorial coming · router config included · avg lift is the mean of the skills shown
Your team’s skills

Author, eval, and share your own — on the same rails.

See skills governance on the registry →

Start with the skills for the agent you’re building.

Browse free — no sign-up. Two lines to instrument your agent.

Start the tutorialor browse the registry