Your GTM Agents Work in Seconds. Your Playbook Still Changes Once a Quarter

Recursive GTM closes that gap. Here’s how it works, and who should control it

Oct 13, 2026

Every quarter, a room full of smart people opens the QBR deck to work out what happened and what to change.

Between those meetings, their agents run around the clock. They draft the outbound and score the leads, then book the meetings and keep the CRM current at 3am.

So the work happens in seconds, and the lessons take a quarter to reach the playbook.

Recursive GTM is the attempt to close that gap: each cycle’s results feed the next one, until the system starts rewriting the playbook it runs on.

I’ve spent a good part of this year writing about loops for coding agents, where a model rewrites its own code and a test suite tells it whether the change worked. GTM runs on the same idea and it’s much harder, because here the test suite is a buying committee and the results show up months later.

On October 15, Hard Skill Exchange is bringing operators from ZoomInfo, Outreach, Gainsight, Dialpad and 1mind online to show how they’re building these loops. AI Practice Sessions: Recursive GTM starts at 9:00am PT, and it’s free to attend.

Systems of Learning

Salesforce launched in 1999, and the CRM has sat at the center of the revenue stack ever since. It’s a system of record: it stores what happened, and the process around it changes when a person rewrites it. Agents added a second layer, a system of action that takes the next step itself. Satya Nadella described where that leads on the BG2 podcast in December 2024: agents are “going to update multiple databases, and all the logic will be in the AI tier, so to speak.”

The third layer is the new one. A system of learning sits on top of both and rewrites the playbook those agents run, with outcomes as the evidence.

The raw material is something CRMs rarely capture: the reason behind each decision. Last December, Foundation Capital's Jaya Gupta and Ashu Garg named that missing layer a context graph, "a living record of decision traces stitched across entities and time so precedent becomes searchable," and their essay calls it AI's trillion-dollar opportunity. A CRM that knows a deal closed at a 20% discount holds a record. A system that also knows a competitor was in that deal, and that the same move won four of the last six similar deals, holds something it can learn from.

How the Loop Works

Recursive GTM runs as a seven-stage loop. The first four stages, where the system builds context and senses change, then decides and acts, are the agent stack plenty of teams already pay for. Stages five to seven close the loop, and the last one, learn and rewrite, is where the system starts editing its own playbook.

Stage six, evaluate what worked, is the hard one in B2B. Tying a deal that closed in March to a sequence that ran in October takes a lot more than counting email opens.

Stage seven is where the recursion lives. Sam Altman, writing about AI tools that help build better AI, called that pattern "a larval version of recursive self-improvement" in The Gentle Singularity in June 2025, and GTM is at the same larval stage today.

Feedback-to-Change Latency

If I could add one number to every GTM dashboard, it would be feedback-to-change latency: the time between an outcome landing and your system behaving differently because of it.

A team that updates its playbook at the QBR runs at roughly 90 days, and weekly sequence tests bring that down to seven. A recursive system aims for hours, account by account.

I’m skeptical of the hours part for anything tied to revenue. Replies and booked meetings can reshape a sequence by lunchtime, while closed deals take a full sales cycle to report back, so the fast loops run on early signals and get checked against revenue months later.

Give two teams the same tools and the same budget, and the one with the shorter latency pulls further ahead every quarter.

Where the Loop Breaks

I’d watch for four failure modes.

It chases the proxy. Point the loop at meetings booked and it learns to fill calendars with people who were always going to pass. That’s Goodhart’s law: once a measure becomes a target, it stops being a good measure. A recursive system just gets there faster.
Marketing starts editing the product. A loop that learns which problems make buyers lean in will eventually hand you roadmap advice. That’s useful right up until the CMO and the CPO start fighting over who decides what the product becomes.
People set the wrong bounds. The scary version is a system that outgrows its limits, and the likelier one is a team that sets the wrong limits on day one. A sales leader paid on bookings and a CS leader paid on retention will configure the same loop in opposite directions, and the loop will faithfully serve whoever wrote the objective.
The data underneath is a mess. This is the one I’d bet breaks first. A loop learns only as well as its context allows, and in most companies that context lives in raw call recordings and Slack threads, plus a CRM someone cleans up on Friday afternoon before the forecast call. Recursion on top of that learns fast, and a lot of what it learns is noise. Your first recursive GTM project is a data project.

Hard Skill Exchange built the free October 15 sessions around the first three.

Before You Switch It On

Start with decision rights. Jeff Bezos drew the line in Amazon’s 2015 letter to shareholders: one-way doors are “consequential and irreversible or nearly irreversible” and deserve slow, careful calls, while the reversible two-way doors, which he labeled Type 2, should move fast.

“Type 2 decisions can and should be made quickly by high judgment individuals or small groups.”
Jeff Bezos, Amazon’s 2015 letter to shareholders

Give the loop the two-way doors and leave the one-way doors with leadership. Write that split down before the loop runs, then hold it to five rules:

▫️ Tie the objective to revenue, with guardrails. Pipeline won is the target, and discount rate and early churn have to hold steady while the loop chases it.

▫️ Keep a holdout. A slice of accounts the loop leaves alone is the only way to tell learning from luck.

▫️ Give it one owner across the revenue team, usually RevOps or a GTM engineer, so the objective reflects the whole P&L instead of one team’s comp plan.

▫️ Log every change it makes to itself. Each entry says what changed and why, and someone reads the log every week.

▫️ Keep a one-click way back to the last approved version of every sequence and score.

This breakdown of why most self-improving AI loops fail shows how a loop drifts while every step looks fine.

Where to start

Pick one loop narrow enough to measure. Outbound messaging for a single segment works well, because replies come back within days and a holdout is easy to set up. Measure feedback-to-change latency before you start and again a month in, and widen the loop’s reach only after the holdout shows lift two months running. If you’re still building the action layer underneath, start with this eight-agent GTM setup and come back.

One test you can run this week: ask your CRO what your GTM system changed last week, and why. If the answer takes more than a minute, start with the log.

AI Practice Sessions, October 15

Hard Skill Exchange built the event around curated 30-minute sessions for the VPs and CXOs who run revenue. The sessions cover how companies move to agent systems built on context graphs, and which operating models fit which kind of company.