Flat 30% off for Singapore 🇸🇬
Book a 30-min call →
21 days · from $1,999 · fixed

Ship It Live · AI agent sprint

An AI agent that does the work, live in 21 days.

Not a chat box — an agent that reads, decides and acts in your systems, with a human able to see and stop everything it does.

Enter to send
4.93★ / 32 reviews 75+ shipped Kickoff in 24h

Quick answer

An AI agent that takes real actions in your own tools can be built and running in 21 days from a fixed $1,999. That covers the task definition, tool integrations, an approval step on anything consequential, run logs and a review interface. An agent with no human checkpoint is not something I will build.

The plan

Day one to day 21, written down.

You get this schedule before you commit, not after. If a phase slips, that is mine to absorb — the price and the date were agreed before anything started.

The scheduleDay 1 → day 21
01Pick the taskDays 1–3
02Design the loopDays 4–9
03BuildDays 10–16
04Shadow modeDays 17–19
05LaunchDays 20–21
  1. 1

    Days 1–3

    Pick the task

    The narrowest valuable loop, its inputs, and what a wrong answer costs. Agents fail when the task is vague, so most of the risk is removed here.

  2. 2

    Days 4–9

    Design the loop

    The decision points, the approval gates, and what the human sees. Designed on paper before anything is wired to a live system.

  3. 3

    Days 10–16

    Build

    Tool integrations, the agent loop, logging and the review interface. Run against real data in a sandbox where nothing it does is permanent.

  4. 4

    Days 17–19

    Shadow mode

    It runs on real work but only proposes — a human approves everything. This week tells you how often it is right before it can be wrong at your expense.

  5. 5

    Days 20–21

    Launch

    Selected actions unlocked with approval still in place on the rest, monitoring on, and a clear kill switch.

What from $1,999 gets you

Everything needed to launch — nothing padding the invoice.

A narrowly defined task the agent is actually good at
Tool integrations — your CRM, sheets, inbox, database, or internal API
An approval step on anything consequential, before it happens
Full run logs: what it saw, what it decided, what it changed
A review interface so a human can inspect, correct and re-run
Guardrails: what it may never touch, and when it must stop and ask
Cost controls and rate limits, so a loop cannot quietly spend thousands
Source code, prompts and configuration yours at handover

Who books this

Built for

Teams with a repetitive process that is judgement-heavy but low-stakes

Operations doing the same research-and-update loop dozens of times a day

Founders who want agents doing work, not demos of agents doing work

Real situations

Five ways this shows up, and what actually fixes it.

Not hypothetical — the shape of the problem before someone books this sprint, and the specific part of the build that resolves it.

Situation 1

Doing the same research-and-update loop dozens of times a day

An operations team repeats a judgement-heavy but fundamentally low-stakes research-and-update cycle many times daily, consuming hours that could go toward higher-value work.

That narrow, well-defined loop gets automated with an approval step on anything consequential — a real agent doing the work, not a demo of one.

Situation 2

Wants agents doing real work, not another chatbot demo

A founder has watched multiple 'AI agent' demos that look impressive but never actually complete a real task inside their own systems.

A narrowly defined task with real tool integrations — CRM, sheets, inbox, database — ships in 21 days, doing the work rather than demonstrating it.

Situation 3

Afraid of an agent doing something consequential unsupervised

A team is interested in AI automation but hesitant to give any system unsupervised access to actions that could cost money or damage a customer relationship.

An approval step on anything consequential is a hard requirement of the build — nothing irreversible happens without a human checkpoint, ever.

Situation 4

No visibility into what an automated system actually did

A previous automation attempt worked, mostly, but left no record of what it saw, decided, or changed, making it impossible to audit after the fact.

Full run logs capture what it saw, what it decided and what it changed, with a review interface so a human can inspect, correct and re-run.

Situation 5

Worried about a runaway process quietly racking up costs

A team considering agent automation is concerned that an unattended loop could spiral in cost without anyone noticing until the bill arrives.

Rate limits, per-run token budgets and a spend cap are part of the build, with the API key held by the client so spend is visible directly.

The honest part

When 21 days is the wrong answer.

A fixed deadline only works when the scope genuinely fits inside it. These are the cases where I will tell you so rather than take the booking.

Fully autonomous agents acting on money, contracts, customers or production infrastructure without a human checkpoint. I will not build that, and you should not want it yet.

Agents whose task nobody can describe precisely. If two people in your team define the process differently, we settle that first.

Replacing an entire role. A good agent takes a defined loop out of a job; anyone selling you the whole job is selling you a rebuild in six months.

Scopefrom $1,999 · 21 days

Inside the 21 days

  • A narrowly defined task the agent is actually good at
  • Tool integrations — your CRM, sheets, inbox, database, or internal API
  • An approval step on anything consequential, before it happens
  • Full run logs: what it saw, what it decided, what it changed
  • A review interface so a human can inspect, correct and re-run
  • Guardrails: what it may never touch, and when it must stop and ask
  • Cost controls and rate limits, so a loop cannot quietly spend thousands
  • Source code, prompts and configuration yours at handover

Quoted as one number before day one. A week running long is mine to absorb.

Outside the line

  • Fully autonomous agents acting on money, contracts, customers or production infrastructure without a human checkpoint. I will not build that, and you should not want it yet.
  • Agents whose task nobody can describe precisely. If two people in your team define the process differently, we settle that first.
  • Replacing an entire role. A good agent takes a defined loop out of a job; anyone selling you the whole job is selling you a rebuild in six months.

Real work, sized and priced separately. You hear it before you book, not at handover.

Questions

Before you book

A chatbot answers; an agent acts. It reads from your systems, decides, and writes back — creating records, sending drafts, updating fields. That is why approval gates and logs matter far more here than in a chatbot.

Task charter (what the agent owns / where it stops)

An agent sprint is for a bounded task with a human on the consequential step — not unsupervised autonomy cosplay.

Write the charter before tools

The agent owns [task] using [inputs / systems]. Done-when looks like [observable outcome]. It must not do [out of bounds] without a named human approval.

If you cannot fill that without saying “it just figures it out,” you do not have an agent yet — you have a demo with permissions.

Belongs in this 21-day sprintBecomes a different project
One reversible first jobOpen-ended “run the company”
Propose → review → act pathSkip-the-human forever
Named approval ownerShared inbox with no owner
Spend ceiling written downUnlimited keys on a shared card

Sibling: if the deliverable is a user-facing product around one job (with uncertainty UX + evals), see /shipitlive/ai-app-in-30-days. If it is a billable SaaS workflow with light AI, /shipitlive/saas-mvp-in-30-days may be the sharper slug.

Shadow mode as a hard gate

The live page already says consequential actions need a human checkpoint. Make the gate checkable.

Definition

Shadow mode: the agent proposes; a human executes — until exit criteria are met. It is not optional theater for week one.

Exit criteria (leave shadow only when these pass)

GatePass looks like
Task charterWritten, shared, no “and then anything else”
Reversible first jobFailure does not burn money or customers by default
Review pathNamed human can approve / reject in one place
Action logEvery run leaves a readable trail (see below)
Kill switchNamed owner who can stop runs

We do not claim unsupervised autonomy after day 21. Soft autonomy: scoped tools, approvals on consequential steps, human ownership of the kill switch.

Action log + spend ceiling

“Run logs” on the page are not a vibe. After each run, a human should be able to answer three questions without digging through provider dashboards alone:

  1. What did it see? (inputs / context summary)
  2. What did it decide? (proposal or action)
  3. What changed? (writes, drafts, skips)

Spend as a done-when (process, not a vendor promise)

  • Hard spend ceiling or per-run budget in the product or provider account you control
  • Alert path when spend spikes
  • What happens when the ceiling hits (pause, queue, message)
  • API key held by you at handoff so spend stays visible

We do not invent runaway-bill dollar figures here. The fear on this page is real; the fix is a written ceiling and a named kill-switch owner — not a scary number we made up.

Tool access (email, CRM, sheets): least privilege for the charter job only. Exact connectors are scoped on the call — no invented integration list in this insert.

Extra questions

Existing FAQ already covers chatbot vs agent, cost, what if it does something wrong, tools, schedule, spend, and ownership.

AI agent vs AI app — which page?

AI agent (this page): bounded task loop, shadow mode, action log, approvals on consequential steps. AI app: user-facing product around one job, with uncertainty UX + evals — /shipitlive/ai-app-in-30-days.

Who publishes / who can approve consequential actions?

A named human (or small named set). Shared “anyone on the team” with no owner is how approvals become theater.

Can it get email / CRM access in the same sprint?

Often yes for the charter job — least privilege, scoped on the call. Broad “company brain with every tool” is out of scope for 21 days.

Is day 21 a guarantee the agent never does the wrong thing?

No. Day 21 means the sprint build, shadow gate, log, spend ceiling, and handoff package are done. Agents still need human judgment on consequential steps — that is why shadow mode and the kill switch are in scope. Day counts are planning, not an SLA on autonomy.

21 days from now, this could be live.

A 30-minute call decides whether the scope fits. If it doesn't, I'll tell you what would.