Product

One goal. One metric. An agent that doesn't stop at done.

Most AI automation finishes a task and waits for the next prompt. AgentWorks gives an agent an outcome to own. It plans the work, runs it on schedule, measures every run and changes its own plan until the metric hits your target.

60-second demo: an agent with the goal "Book more sales demos". Illustrative data.
  1. Set the goal

    An outcome in plain words, the metric that proves it, and the target.

  2. Run

    Agents plan the steps, connect your tools and run on schedule.

  3. Measure

    Every run records what it did and what it moved.

  4. Auto-improve

    It fixes what broke, drops what didn't work and tries what's next.

Goals

Say what you want. Pick the number that proves it.

A goal is the outcome, a primary metric with a target, a few supporting metrics, and the rules that must stay true while the agent works. Describe it in chat and AgentWorks sets it up with you.

  • One primary metric. The number that decides whether the goal is met.
  • Supporting metrics. The signals that explain why it moved, like reply time or open rate.
  • What must stay true. Guardrails the agent can't trade away, like "never email the same person twice a day".

What we're working toward

Turn more inbound signups into booked demos.

Primary metricDemos booked per week
6target 5

SupportingFirst reply time · Reply rate · Show-up rate

Must stay trueMax 2 emails per person a week · Never promise pricing

Measure

Progress, not activity. And never a made-up number.

Every run leaves evidence: what it did, what it cost, and what the metric did next. If a number is missing or out of date, the goal says so instead of guessing.

  • Trend, not a snapshot. Every measurement is dated, so you see the direction.
  • Stale data flagged. "Measurement stale" beats a confident wrong number.
  • Cost per goal. What each goal costs to run, per run and per model.

Auto-improve

It finds what would move the goal. Then it does it.

After runs, AgentWorks reviews the evidence against your goal. It repairs broken steps, drops ideas that didn't work, does the work nobody was doing, and comes back later to check whether it helped.

  • Fixes. A step failed or a login expired: it repairs it and re-runs.
  • Did for you. Each change says what it should move and when it will check.
  • Focus areas. Tell it where to look first. Your goals and rules always win.
  • Challenges your rules. If a rule is costing the goal, it asks. The rule stays until you answer.

Autonomy

You decide how far it goes. Turn it up as you trust it.

Set it per goal. Anything that reaches a customer can always require your approval, whatever the level.

  1. Level 1

    Ask first

    Proposes every step and waits for your yes. Good for week one.

  2. Level 2

    Run steps

    Runs the workflow on its own. Asks before anything goes out or the workflow changes.

  3. Level 3

    Edit workflow

    Also rewrites its own steps when the evidence says a change will move the goal.

  4. Level 4

    Full

    Runs, changes and ships within the rules you set. You review the log, not each step.

Layer 1 · Agents

Real agents, not a model behind a wrapper.

Most AI tools call a model and bolt their own loop around it. AgentWorks runs each vendor's own agent, with the model and the harness its makers built and tuned together, so you get everything the agent can do.

Your goalDemos booked per week · target 5AgentWorks plans, schedules, measures and improves
Sandbox · only the folders this workflow was granted
Native agent in a live terminal
  • Claude Code
  • Codex
  • Cursor
  • Pi
  • Muse
Model + harness from the company that built it · watch it live, step in anytime
MCP serversYour tools and AgentWorks' own
Shell & CLIsgit, gh, cloud CLIs, scripts
BrowserIts own persistent, isolated session
  • The full agent, not just the model

    Claude Code, Codex, Cursor, Pi and Muse run as themselves in a live terminal session, with the planning and reasoning their makers built. That is also why they run on the subscription you already pay for.

  • Your tools, governed

    Every agent gets your MCP servers, AgentWorks' own tools, a shell and a browser, granted per workflow. Writes go through AgentWorks' tools, so every change is checked and logged.

  • Walled off from everything else

    An OS-enforced sandbox (Landlock on Linux, sandbox-exec on macOS) limits each agent to its workflow's folders, with separate read and write access. Your other files, other workflows and machine secrets stay out of reach.

Layer 2 · Crews

A specialist with a memory, a desk and a phone.

A crew is an agent turned into a long-lived teammate. It keeps its own workspace, skills and memory, has its own browser, answers in Slack or WhatsApp, runs on schedules, and can call other crews when a job needs a different specialist.

Slack WhatsApp ChatGPT Claude & Cowork Any MCP client
  • AgentAny of the five native agents from Layer 1, inside its sandbox.
  • SkillsPicks from your shared skill library and writes its own as it learns the job.
  • MemoryA written memory of your customers, preferences and decisions, plus one ongoing conversation that survives restarts.
  • WorkspaceIts own files, database and live dashboard.
  • BrowserIts own managed browser for any web app, with a live view you can watch.
  • ChannelsTwo-way Slack and WhatsApp. Gmail for updates and for mailbox work you grant.
  • Crew-to-crewCrews call each other like colleagues: ask a question or call a typed function, each in its own private thread.
  • Triggers & schedulesRuns on schedules, signed webhooks, function calls or a message in chat.

Layer 3 · Workflows & goals

A full plan, not a prompt. Code where it should be, agents where it matters.

A workflow is the whole operation written down: pipelines of steps, some fixed and deterministic, some handed to an agent or a crew, all measured against one goal. Each run adds to what it knows.

Book more sales demosGoal: 5 demos a week

Pipeline · New signups

  1. DeterministicPull new signupssaved script
  2. AgentResearch each companyClaude Code
  3. DeterministicRoute by company sizerule, no model
  4. CrewSage writes the emailcrew step
  5. ApprovalYou approve the sendhuman step

Pipeline · Follow-ups

  1. DeterministicFind quiet signupssaved script
  2. AgentDraft the check-inCodex
  3. DeterministicLog to the goalmetric update
DeterministicAgentCrewHuman approval
  • Deterministic where it should be

    Routing is plain code that reads a file, never a model guessing. Steps that repeat can save a script that runs first next time, so a solved problem stays solved and costs nothing.

  • Agentic where it matters

    Research, judgment and writing go to an agent step or to a crew, with the right model for each step and a human approval wherever you want one.

  • Learnings & knowledge base

    Every run writes what it learned into a shared skill, and each workflow keeps a knowledge base that other workflows can read. Run fifty knows what run one had to discover.

Bring your own AI

Use the AI plan you already pay for.

AgentWorks runs on your Claude, ChatGPT, Gemini or Cursor subscription. No credits, no per-task fees, no marked-up tokens.

  1. Connect your planSign in once with Claude Code, Codex, Cursor, Pi or Muse, using your subscription or that CLI's API key.
  2. Match the model to the jobYour strongest model for judgment, a cheaper one for routine runs.
  3. Keep working at the limitWhen a plan hits its limit, work pauses and resumes at the reset, or moves to another plan you've connected.

Connected AI plans3 plans · 4 teammates

Claude MaxJudgment, planning, customer-facing draftsIn use
ChatGPT ProResearch, reports, routine runsIn use
GeminiTakes over when a plan hits its limitStandby

Under the hood

Built on the tools you already trust.

  • Your AI plan

    Runs each vendor's own coding agent (Claude Code, Codex, Cursor, Pi or Muse) on the subscription you already have, or on that agent's API key.

  • Its own browser

    A persistent, isolated browser per workflow, so agents can work in any web app you use.

  • MCP and APIs

    Connect any MCP server or API. Tools are granted per workflow, not globally.

  • Secrets vault

    Encrypted, injected only at run time, never shown in chat or logs.

  • Sandboxed

    OS-enforced sandbox per agent: Landlock on Linux, sandbox-exec on macOS.

  • Full run log

    Every step, tool call, decision and cost, kept per run.

FAQ

How it works, answered.

What makes a good goal?

One outcome you care about and one number that proves it, with a target. "Book 5 demos a week" works. "Do more marketing" doesn't, because nothing can tell whether it happened.

What if the metric can't be measured automatically?

AgentWorks says so. A missing or stale measurement is flagged on the goal instead of guessed, and the agent asks you how to get the number.

Can it change my workflow without asking?

Only if you let it. At the default autonomy level it runs steps on its own and asks before posting, sending or editing the workflow. You can move it up or down at any time.

What's the difference between a Goal and a Crew teammate?

A Goal owns an outcome and keeps working on its own schedule. A Crew teammate is who you talk to: you hand it a job in Slack or WhatsApp, or ask how a goal is doing.

Can I use it from ChatGPT or Claude?

Yes. AgentWorks includes an MCP server. Connect it to ChatGPT, Claude, Cowork or any MCP client to check goals, read reports and start runs from that chat. Hosted AI apps connect to your AgentWorks server with a sign-in link.

Does it work with the tools I already use?

Yes. Agents use MCP servers, APIs and their own browser, so anything you can do in a web app, they can do too, with credentials kept in the vault.

Pick one goal. Watch the number move.

Start from a premade agent or describe your own goal. Ten minutes to set up, on the AI plan you already have.