You Can Hire an AI Team. Who Writes the Spec?

Here are two shapes AI help for a small business can take. One sells you a team: AI employees with names and job titles, working in your inbox, your social accounts and your phone line. The other starts from a written spec for the thing you want built, splits the work into tasks, organizes the workers around shared contracts, and checks each piece before it counts as done.

TresPies works the second way. Cruz Morales, who runs TresPies, says its contract work is "built off of specifications and shared contracts." Below: what each shape shares, what that lets you check, and what to ask before you pay for either.

What an AI team looks like

Marblism is a clear example of the first shape. It sells seven AI employees to businesses: an executive assistant, a social media manager, a lead generator, a website builder, an SEO writer, a receptionist that answers calls around the clock, and a legal assistant. One plan covers all seven, starting at $24 a month billed yearly, for 50 hours of AI work a month. Marblism says the team "shows up every single day."

The employees share some things. Marblism's help center describes a "Brain", which it calls "the company knowledge your AI employees share": web pages, documents and files you add. For each item, you choose which employees can see it. Employees can also "talk to each other and collaborate on tasks." In the help center's example, you ask the social media manager for a campaign, and it coordinates with the SEO writer to pull insights from the writer's latest blog posts.

One thing is kept per employee. The help center says "what one agent remembers stays with that agent"; to share something with the whole team, you add it to the Brain. And Marblism's trust center says sending email, posting publicly, signing or returning contracts, and outbound calls "all wait for your explicit approval."

What contract work looks like

Contract work is organized around the thing being built. It starts with a written spec: what the finished thing must do and how you will know it works. Then come contracts between the pieces: what each part receives, what it hands back, and in what form. Each task gets a worker with a clear job and clear limits. Nothing counts as finished on a worker's word. TresPies follows the rule in the Orchestration Guide from Dojo Genesis, the developer brand Cruz founded: "Agent 'done' is not done until independently verified." Cruz calls this protocol, "the system as a whole, a series of checks and hooks." (A hook is a check that runs on its own at a set point in the work.)

Anthropic, the company that makes Claude, describes the same need in a research system it built out of many agents. Its lead agent has to give each worker "an objective, an output format, guidance on the tools and sources to use, and clear task boundaries." Without that, Anthropic writes, "agents duplicate work, leave gaps, or fail to find necessary information."

How the work is divided matters too. In one experiment described on Anthropic's site, agents split by software role (a planner, an implementer, a tester and a reviewer) "spent more tokens on coordination than on actual work." Anthropic's advice is to divide work by context instead, so that "an agent handling a feature should also handle its tests." That experiment was about coding agents, and Anthropic describes it in one sentence. It says nothing direct about a product like Marblism.

The tools, in public view

TresPies builds with Claude Code, Anthropic's coding tool. Here is what three of its features do, as Anthropic documents them.

Feature	What it does
Projects	One ongoing conversation where Claude acts as coordinator and starts a separate thread for each task. Threads, usually cloud sessions, run in parallel and report back when they finish. Cloud threads share the project's code, instructions and memory. In public beta on Pro and Max plans.
Workflows	A script decides which agents run, in what order, and which run side by side. An agent can be given a required output format, and the step fails if its answer still doesn't match after five tries.
Artifacts	Results are published as a private web page you can share and update in place, and each publish is saved as a version. People you share it with inside an organization (Team and Enterprise plans) can comment on it.

Sources: Projects, Workflows and Artifacts in the Claude Code docs, read 2026-10-25.

Why the spec carries the work

Cruz describes the spec doing a different job on different kinds of work. On backend work (the systems behind an app), shared specs and contracts "organize and distribute agentic workers." On front-end work (the screens people use), specs and memories from the shared workspace point workers to the resources that inform the outcome, based on the client's intent. Research generated along the way "also informs outcomes."

Anthropic's guidance supports the backend point: "With a well-defined API contract," it writes, "frontend and backend work can proceed in parallel."

Here is one difference in what each shape shares. Marblism's employees share a Brain of company knowledge, and you approve sensitive actions before they go out. Contract work shares a written spec of what done means. You can read it and hold the finished work up against it. TresPies gives its clients a copy of the spec.

What it costs, and where it falls short

Splitting work across many agents costs more, and it does not suit every job. Here are four limits from the public record.

It uses more. In Anthropic's testing, multi-agent setups "typically use 3-10x more tokens than single-agent approaches for equivalent tasks." Claude Code's own documentation says a workflow "can use meaningfully more tokens than working through the same task in conversation."
Not all work splits. Anthropic notes that "most coding tasks involve fewer truly parallelizable tasks than research." A study from Google Research, Google DeepMind and MIT (first posted December 2025, revised April 2026) found multi-agent setups did 39% to 70% worse than a single agent on step-by-step planning, depending on how the agents were organized, and 2% to 15% worse on a standard coding test.
Shared files collide. "Two teammates editing the same file leads to overwrites," Claude Code's documentation says, so each worker should own its own files. Saying who owns what is one job a contract does.
Checks can be shallow. Anthropic calls the most significant failure of a checking agent "marking outputs as passing without thorough testing."

So the case for contract work rests on the written target each piece is built to, and the checks it has to pass. Sometimes one agent is the better choice: in Anthropic's comparisons across several enterprise deployments, a single agent with skills "frequently outperformed" designs that split the work across subagents on quality.

Cruz's answer to this record, from his own work: "Well governed and researched dispatches [tasks handed to agents] save both time and produce meaningfully better results at a higher consistency." He sees this "across coding" and beyond, and says it "has to do with protocol," which "solves problems like parallel writes and routes simpler tasks away from complex protocols."

What to ask before you pay for AI work

Ask	Why it matters
Is there a written spec, and do I get a copy?	It is how you check the finished work against what you asked for.
What does each piece hand to the next?	Clear hand-offs keep pieces from overlapping or leaving gaps.
Who checks each piece, and how?	A worker reporting its own job as done is not a check.
Where does the finished work live?	If it sits in the provider's accounts, ask what happens when you leave.
How does the price change if the job grows?	Some tools meter hours of AI work, others tokens. In Anthropic's testing, multi-agent setups typically used 3 to 10 times more tokens than a single agent on the same task.

A first step

Write one page on what the thing you want built must do and how you will know it works. That page is the start of a spec, whoever builds it. If you want TresPies to build from it, you can book a 30-minute conversation with TresPies and bring the page. TresPies clients get a copy of the spec their work is built from.