GTM workflows for an AI agent team: shared context and a handoff template

Why do GTM workflows break when several AI agents share the work?

Because nobody wrote down who owns each step, what it reads and what it hands back. A research agent finds three facts about an account. The drafting agent can't see them, so it searches again or guesses. The draft moves on with no proof behind its claims.

OXYGEN co-founder Tim Scheuer met both halves of this problem. In September 2026 his 40+ skills were "1 big pile". One chat and one job per agent made it "so much easier for me to assign specific things to specific agents". The quality stayed hard: "Where I still struggle is the quality of the output".

So an AI agent team needs three things that roles alone don't give it:

• One owner per step, so two agents never research the same account.
• Evidence that travels with the work, so every claim in a draft has a source.
• A check before each handoff, so weak work stops instead of moving on.

Sources: Tim Scheuer on LinkedIn: 8 GTM employees on Grok Bot (September 2026)

How does Tim Scheuer split GTM work between his agents?

Into five layers, with goals at the top and the GTM platform at the bottom.

He split the work further over time: outbound went from 1 agent to "5 specialised ones", and group chats let them talk. When he tested Grok Bot, "the quality wasn't there" for copy, design and launch videos. Now "Grok Bot orchestrates. Claude Code and Codex do the work with my skills."

This is one founder's setup from September and October 2026. OXYGEN does not ship his 13 roles as ready-made agents. His October 2026 post lists the five layers, and each plays a different part in a handoff:

• "Me. Goals and approvals." The person sets goals and says yes or no to live work.
• "My Chief of Staff routes the work." One of his 13 agents routes; each of the others owns one job.
• "Any agent can send them a unit of work." Claude Code and Codex do the work.
• "Every agent reads from it and writes back to it." His second brain, synced to GitHub, is the shared reference.
• "My agents run my entire GTM in that platform." His GTM stack lives in OXYGEN.

Sources: Tim Scheuer on LinkedIn: the 5-layer org chart (October 2026), Tim Scheuer on LinkedIn: 8 GTM employees on Grok Bot (September 2026)

What goes in a handoff between two GTM agents?

Eight fields, filled in once per handoff and sent with the work. The posts name the layers but not what passes between them, so this card fills that gap.

The habits come from the six prompts in our GTM skills kit. They keep facts apart from guesses, leave unsupported details unknown and wait for approval before paid or live steps. Copy the card from the setup steps below. Each field answers a question the next agent would otherwise guess at:

• Input: what you hand over, such as one account row, and the goal it serves.
• Owner: the one agent or person who does this step. Two owners means nobody owns it.
• Shared reference: the pages to read first, such as the ICP (ideal customer profile), offer and voice guide.
• Expected output: the artifact and its shape, such as a five-field brief or one email draft.
• Evidence: a source for every claim, and each missing fact written as unknown.
• Review or approval: what a person must see, and which actions need their yes, such as a send.
• Stop condition: when to hand back instead of guessing, such as a missing domain.
• Next owner: who gets the output, and where it goes if it fails their check.

Sources: OXYGEN GTM skills kit: start with a useful job

What does one handoff from account research to an outbound draft look like?

Here is one, from the goal to a draft a person approves. Acme Software and every fact about it are made up for this page. The goal: a first email to Acme's VP Sales.

The router's card sends Acme's row to the research agent, with the ICP and offer pages as the shared reference. It asks for a brief and stops the step if the domain is missing or Acme is a current customer. The brief that comes back, also made up:

• Fit: possible. Fact 1: it sells software to companies and has 140 staff (company site, LinkedIn page).
• Fact 2: it is hiring two SDRs, the sales development reps who book first meetings (careers page).
• Fact 3: its job posts name HubSpot as its CRM, where customer records live (job post).
• Unknown: contract value, and who owns the outbound tools.
• Angle: new reps need a researched first list. Next owner: the drafting agent.

What happens when a handoff fails its quality check?

It goes back to the step that can fix it, with the gap named. In the example, the first draft opened with "Congrats on the Series B". The brief had no funding fact, so the reviewer failed the draft and listed that line.

The drafting agent could cut the line or ask research one question. It asked. Research found no funding news in its sources and marked funding unknown. The second draft opened "You're hiring two SDRs [Fact 2]" and passed. Two failed reviews in a row go to a person, never to a third rewrite.

Use the same branches for every failure:

• A claim with no source: back to drafting to cut it, or to research for one fact.
• A missing input, such as no domain: research stops and hands back to the router.
• A stale reference, such as an old offer page: stop, and ask the page's owner to update it.
• Two failed reviews on one piece of work: a person decides what happens next.

How do agents share context without repeating research?

Keep one shared reference that every agent reads before it starts and writes back to when it learns something. That is Tim's fourth layer: his second brain, "synced to GitHub. Every agent reads from it and writes back to it."

Split it in two. Reference pages hold what rarely changes, such as your ICP, offer, voice and do-not-contact rules, and a person approves changes to them. A claim ledger holds what agents find about each account, so the next agent reads it instead of searching again. Tim's team described how it built its company-wide second brain.

Give the ledger one row per claim. Here is the example's hiring fact as a row:

• Claim: Acme is hiring two SDRs.
• Source: the careers page, with the date it was checked.
• Added by: the research agent.
• Status: verified, unknown or wrong. A draft may use verified claims only.
• Used in: the draft line that relies on it, so a reviewer can trace it back.

Sources: Tim Scheuer on LinkedIn: the 5-layer org chart (October 2026)

Where should a person review before anything goes live?

At every step that leaves the workspace or spends real money: a send, a sequence start, a post or a large paid lookup. A sequence start launches an automated series of outreach messages. Tim keeps that layer for himself: "Me. Goals and approvals." For his outbound team, "Nothing goes out without my yes."

Write the boundary into the card, so every agent knows where it stops. Approve a concrete scope: the audience, the copy, the schedule, a send cap and the cost. Our kit's outreach prompt ends the same way: "Keep everything unlaunched until I approve the concrete scope and budget".

Sources: Tim Scheuer on LinkedIn: 8 GTM employees on Grok Bot (September 2026), Tim Scheuer on LinkedIn: the 5-layer org chart (October 2026), OXYGEN GTM skills kit: start with a useful job

Which GTM workflows should an agent team run first?

Start with the six jobs in our GTM skills kit, and chain them so each job's output is the next one's input.

The kit runs inside OXYGEN, or in Claude Code and Codex with an OXYGEN account. The skills have no separate fee; data, enrichment (filling in missing emails and company facts) and sending use your plan's credits. The kit says what a good result holds:

• Decide what to fix first: a sourced snapshot, one proposed bottleneck and the question that could change it.
• Build a target-account list: fit criteria, a preview, the proposed scope and a priced next step.
• Find buyers at those accounts: likely buyer roles, kept apart from people you have verified.
• Research and qualify accounts: evidence behind every grade, with unknowns kept apart. Our lead qualification workflow covers this step.
• Prioritize warm opportunities: fit and relationship evidence and a next action. Engagement alone is not buying intent.
• Prepare an outreach pilot: a plan you can review, with nothing running yet.

Sources: OXYGEN GTM skills kit: start with a useful job, OXYGEN GTM skills kit: what it needs and costs

How do you run these handoffs in OXYGEN?

Give each Oxygent one job, and keep the shared reference in Knowledge. An Oxygent is OXYGEN's in-app AI agent: you add one, name it and tell it its job in your chat with it. None comes with a preset role, so the router, researcher and reviewer here are jobs you write.

One Oxygent can message another, and the answer comes back to the one that asked. It can't message an Oxygent that asked it for the work, and one exchange holds at most 12 messages. So let the router hand out each step and collect each answer. A group chat holds two to six Oxygents with you.

You start the work and set its limits. In the composer's approval menu, Ask for approval makes sends, sequence starts and larger paid steps wait for an approval card. A Workspace Oxygent's chat starts on Auto-approve, and each person's choice covers only their own messages. So everyone who uses an Oxygent that sends or starts sequences should pick Ask for approval in its chat. Where the rest lives, and what it costs:

• Shared reference: Knowledge holds your company profile, ICP and offer. Name the pages each Oxygent should read first.
• Voice: pinned voice, brand and positioning pages ground AI-written copy, and a change to one waits as a proposal for approval.
• Team memory: what the team asks an Oxygent to remember is a Knowledge page that every chat with it reads.
• Cost: an Oxygent's model use bills workspace credits unless you connect your own AI model key, and paid steps bill as usual.

What this template does not do

• The roles here are Tim's personal setup and a made-up example. OXYGEN does not ship them as ready-made agents.
• A handoff card checks evidence. It cannot make copy good: in September, Tim still went back to other models for copy and design.
• We have not measured the time or money this template saves.