▸ eve/agent-library

a builder's field guide · compiled 2026-07-25

The one idea

An agent is a
directory of files.

Eve owns the agent loop the way Next.js owns routing. You describe what the agent does — a file per capability — and the framework handles durability, sandboxing, approvals, channels, and evals. This library teaches you to do that well.

● eve-scout agent — built & verified 3 tools · 2 skills · 1 schedule

agent/

agent/
  agent.ts          # the model it runs on
  instructions.md   # who it is
  tools/            # what it can do
  skills/           # what it knows
  subagents/        # who it delegates to
  channels/         # where it lives
  schedules/        # when it acts alone

The five building blocks

What each file is for

Keep instructions tiny and identity-only. Push procedures into on-demand skills. That single discipline is the difference between a good agent and a bloated one.

instructions.mdAlways-on identity, tone, standing rules. Never procedures.

tools/*.tsTyped actions (Zod). Filename = tool name. Runs in your app runtime.

skills/*.mdLoad-on-demand know-how. Only the description is seen until loaded.

subagents/*/A specialist with a clean context window. Split only for a reason.

channels/*.tsSlack, Discord, HTTP, web. One agent, every surface.

schedules/*.tsA cron file. The agent acts on its own clock.

The catalog · study these

Vercel runs its company on 100+ agents

Each takes on a role, not a task. The lesson in each is worth more than the description.

Agent	Role	Metric	The lesson
d0	Data analyst in Slack	30k Q/mo	Every query runs as the asker — security is per-user, not per-agent.
Lead Agent	Autonomous SDR	$5k/yr → 32×	Agent-as-employee. Narrow + always-on beats broad + idle. One part-time engineer.
Athena	RevOps cockpit	6 wks · ~2× pipeline	Non-engineers ship revenue tools when the plumbing is gone.
Vertex	Support engineer	92% solved	Escalation is a feature — hand humans the hard 8%.
draft0	Content reviewer	pre-human	Agents as pre-processors: do the obvious work before a person looks.
V	Router / triage	1 front door	A fleet needs a router so a hundred agents feel like one.

Six archetypes

Name the pattern before you build

Every agent above is an instance of one of these. So is anything you'll build.

read

The Analyst

Q&A over a warehouse/CRM. Per-user scoping, exact numbers, a bounded read-only query tool.

write · cron

The Autonomous Worker

One narrow job, done continuously. Tight scope, idempotent actions, the playbook of your best human.

retrieve

The Support Deflector

Answer from a corpus, escalate the rest. Retrieval + a confidence threshold + a clean handoff.

gate

The Reviewer

Do the obvious work before a human; never publish directly. Structured output + a validator gate.

route

The Router

Classify a task and delegate to the right specialist. Cheap classifier, subagents as targets, fallback.

research

The Scout

Gather, vet, and file knowledge on a cadence. ← eve-scout is this one.

The playbook · non-negotiables

Seven rules that separate good agents from bloated ones

▸Identity in instructions, procedures in skills. A system prompt full of playbooks is the most common mistake.
▸Bound every tool output. Slice rows, cap size, flag truncation. Unbounded returns cause "context rot" — 19 good tools beat 46.
▸Make side effects idempotent or gate them. An interrupted step re-runs on resume — ungated charges & emails double-fire.
▸Write a skill's description as the triggering task. "Use when…" — it's the only routing signal the model sees before loading.
▸Single agent first. Orchestrate subagents only for genuinely parallel work — multi-agent costs 3–10× the tokens.
▸Evals are the deploy gate. Write them before the feature; run eve eval --strict in CI. Grade the trajectory, not just the answer.
▸Security is containment, not prevention. Type-and-gate every tool, approve the irreversible, sandbox untrusted code.

Agents as a business

Where the leverage is

Vendor-reported figures — directional, not audited. The pattern holds: narrow, always-on, well-measured jobs with a human on the irreversible actions.

32×

Return on an autonomous SDR costing ~$5k/yr

vercel · Lead Agent

92%

Support tickets solved with no human

vercel · Vertex

~317%

Median AI-agent ROI · ~5-month payback

market · Articsledge

29%

Of Vercel deployments now triggered by agents (was <3%)

vercel · eve blog

You asked what to ask

Questions that pressure-test any agent

The full set of 30 lives in 06-questions-to-ask/. A starting eight:

What is the single job this agent does — in one sentence?

What does one wrong action cost? That decides the approval gates.

Could a skill do this instead of a subagent? (Default yes.)

Which tool outputs are unbounded, and which actions non-idempotent?

What's the eval that proves it works — written before the feature?

What's my prompt-injection blast radius?

What role does this replace, and how is that role measured today?

If I add one capability this week, what has the highest ROI?

An agent is adirectory of files.