Growth teams used to wait weeks for an engineer to set up an A/B test or change a landing page. Terminal-based coding agents such as Claude Code have removed much of that wait: with the right engineering permissions, a marketer can now build a feature end to end and ship it. The new problem is focus: an agent without structure drifts, and solo founders end up chasing technical side projects instead of revenue. A former head of growth at ElevenLabs, who helped steer the company through hypergrowth, built an open-source framework called Agentic Growth OS to keep Claude Code pointed at business targets.
Why a terminal agent needs guardrails
Web chat assistants tend to produce surface-level text. An agent in the terminal reads the files in a project folder, pulls metrics, runs scripts and edits the assets that actually ship. When anyone can build a minimum viable product in a morning, the bigger risk is no longer slow coding but efficiently building something nobody wants.
Agentic Growth OS is a harness, meaning the instructions, rules and tool connections that wrap around the model. It lives in a public GitHub repository, installs with two Claude Code plugin marketplace commands and also runs on agents such as OpenAI Codex and OpenClaw. It rests on four pillars:
- Goal cadences that link every task to annual and quarterly targets
- Experiment discipline that governs how tests start, run and end
- Specialist agent roles, including a chief of staff, a search engine optimization (SEO) analyst, a product lead who drafts product requirement documents (PRDs) and an adversarial review panel
- Tool connections through APIs and the Model Context Protocol (MCP), a standard that lets AI models reach outside tools and data
A goal cascade from annual mission to daily standup
Each task is weighed against the goals above it to stop scope creep.
| Cadence | What happens |
|---|---|
| Annual mission | A file records the company’s one objective for the year |
| Quarterly OKRs | Objectives and key results set measurable outcomes, such as pipeline volume or activation rate, that every experiment is scored against |
| Monthly retro | The agent compares Stripe revenue and Attio or PostHog engagement data with quarterly baselines to check the pace toward the annual goal |
| Weekly retro | A one-hour spoken review, dictated with a tool such as Wispr Flow, is checked against conversion data for gaps between gut feeling and results |
| Daily standup | Five minutes on yesterday’s results and the one deliverable due by midnight |
The framework’s author calls the daily standup the least valuable and often skips it, but has run weekly retros 10 to 15 weeks in a row, since a weekly cadence yields enough data to reveal patterns. For a one-person operation, the weekly review looks like the place to start.
Onboarding starts with a growth setup command that scans the project folder and sources such as Asana and Google Docs for existing goals. On a sample workspace for a fictional product marketer, it flagged stale OKRs from the second quarter of 2025, a missing monthly target and a pricing-page test left unanalyzed for 14 days. A five-minute rapid onboarding or a deeper session of one to two hours then fills the gaps.
▲ From annual mission to daily tasks
Session hooks that push back
The plugin customizes CLAUDE.md, the instruction file Claude Code reads at the start of each session. A session hook, a script that runs automatically at a set moment, then prints a summary at every launch: an annual mission such as £1 million in annual recurring revenue (ARR) by December 31, a quarterly goal of lifting lead-to-customer conversion from 10% to 20%, a rule of keeping 10 conversion tests running, this week’s priorities, the day’s to-do list and tests that have gone stale.
Claude then behaves like an internal growth lead. It asks whether a task is a 10x bet, and if you start building a 3D game with the Astro framework, it points back to the mission. Connected to the Todoist task manager and GitHub, Claude went further: after two or three retros, it noticed outreach tasks sitting untouched for weeks while the commit history showed refactoring with no growth rationale. It caught procrastination disguised as engineering.
Experiments need a baseline and a kill switch
No test launches without a pre-registered baseline, a clear success metric and a fixed measurement period, and A/B tests also get a power check on sample size. Every test must reach a clean conclusion and land in a shared log, so lessons compound. Without one, new hires can end up proposing tests that already failed, as happened at the author’s first startup. Each experiment follows a template:
| Field | What it captures |
|---|---|
| Hypothesis and key result | What should change, and which key result it serves |
| Audience and treatment | Who sees the test and which variant |
| Baseline and window | The starting value and how long to measure |
| Confounders | Other changes that could muddy the result |
| Kill criteria | The metric drop or bug that ends the test immediately |
| Post-mortem | What the team learned |
Kill criteria fix the stopping point in advance, so a test cannot drag on while a funnel degrades. Ask the agent for an exit-popup test on a landing page, and it first consults the template and playbooks, then asks for the missing baseline conversion rate, weekly traffic and the goal the test supports. Baselines should come straight from an analytics tool such as PostHog, Mixpanel or Amplitude, not be copied by hand. And easy tooling is no reason to run tests on low-traffic or unimportant pages.
Make agents argue with the plan
The framework’s PRD skill borrows Toyota’s 5 Whys, a root-cause method that asks “why” repeatedly, and turns it on whoever proposes a feature. Suggest a landing page meant to win search traffic, and the agent checks keyword volume to see whether it would move a core metric, often arguing against building it when the numbers are thin. Approved PRDs get task checkboxes, so the agent can resume after its context window, the amount of text a model can hold at once, resets or a session dies.
Before code goes out, before payment systems change and before a strategy document is final, the author runs a skill called multi-agent-refine. It spawns three to ten subagents, helpers launched by the main agent, that read the same document through different lenses such as technical correctness, missing considerations, engineering simplicity and operational feasibility. A particularly useful critic asks what is missing, because authors rarely see their own unstated assumptions. Critics with competing criteria, the reasoning goes, catch far more than one agent reviewing its own work.
When the skill ran against a growth OKR document, four critics, covering measurability, feasibility, strategic alignment and gaps, flagged problems including:
- Second-quarter retro data that conflicted with third-quarter goals
- Trial-to-paid targets with no measurement timeline
- An activation baseline drawn from 888 self-serve users, never updated for a trial-to-paid rate of 3.12%
The main agent sorted the findings into P0, P1 and P2 and asked whether to close out the quarter, reset the baselines or patch the document. Watch the cost: launching 20 Opus subagents at once sent the author straight into his usage limits, so he recommends faster, cheaper models such as Sonnet for these panels.
▲ Multi-agent critique of a plan
Connect tools that agents can operate
The stack sticks to services with a solid command-line interface (CLI) or a well-documented API, so agents can read and change data directly rather than rely on hand-reported numbers.
| Tool | Role in the stack |
|---|---|
| PostHog | Product analytics and experiments |
| Stripe | Payments and revenue |
| Attio | Customer relationship management |
| DataForSEO | Keyword and search data |
| Firecrawl | Converts web pages into clean Markdown |
| Resend | Transactional email |
For SEO, the author adapted a public SEO audit skill with more than 206,000 installs, and keyword data comes from the DataForSEO API at cents per query rather than from subscription suites such as Ahrefs or Semrush. The /seo-audit command checks header tags, robots.txt, sitemaps and link structure, and runs the Google Lighthouse CLI to measure page speed, which affects rankings and conversion. On a newsletter site with a large paywalled archive, it found 43% of posts locked away from search engines, 448 posts without a meta description and 130 titles longer than 65 characters. Work that once meant two to three weeks of waiting on an agency, plus weeks of developer time, now takes a repository agent a couple of hours to fix and test.
PostHog is the first tool the author tells growth teams to connect, thanks to a generous free tier and full MCP, API and CLI support. Claude Code built every dashboard, chart and experiment he uses through the API, including a test of one-column versus two-column layouts for a page’s top banner. Three safeguards keep the tracking clean:
- Repository access. Growth marketers should seek at least read-only GitHub access, so agents can check when an event fires without pinging engineers on Slack.
- A CI gate. A continuous integration check in GitHub Actions blocks any pull request whose new code lacks PostHog tracking events.
- Event documentation. After an outside agency misread data because event names had no descriptions, Claude Code and a multi-agent review annotated more than 350 event definitions in about two hours, work that once took a senior product manager weeks.
Keep a human in the driver’s seat
In the author’s judgment, agents that make money for a business with no human involved are mostly a fantasy spread on social media such as X. Unsupervised agents tend to burn tokens, the units of text that AI models process and bill by, on low-value loops or return shallow summaries. He keeps a person in the driver’s seat and hands agents well-scoped research or coding tasks of one to two hours. Scheduled routines already pay off, such as a Sunday-night job that tests the site and emails a warning if performance slips.
Code needs a second set of eyes. Claude Code can state wrong conclusions with full confidence and introduce subtle bugs even when tests exist, so the author pairs it with Cubic, an AI code reviewer for pull requests, and Sentry to flag errors as they happen. He sees networks of agents running and checking experiments overnight as the next frontier, but admits reliable 24/7 operation is still being worked out. In his view, as agents take over execution, the growth marketer’s job shifts toward understanding customers well enough to aim them at the right experiments.
Let the agent rewrite its own instructions
The biggest mistake, by the author’s account, is never asking the agent to improve its own instructions. The fix is to ask Claude how its system instructions could be better, given recent sessions and Anthropic’s latest documentation, and let it rewrite its CLAUDE.md file. A recruiting startup that had its agent critique its own system prompt got a 10-step improvement checklist back, and Agentic Growth OS runs a weekly routine that reviews past sessions, suggests new skills and prunes stale instructions. This is prompt tuning through feedback, not self-improvement in the artificial general intelligence sense, and it relies on file-based tools such as Claude Code and Cowork, which, unlike cloud chat, can edit their own configuration.
What a small team can do this week
The lesson appears to be less about giving Claude Code more work than about giving it structure: goals to answer to, experiments to close out and critics that test plans before they ship. A person sets the direction; the agent handles scoped execution and checks.
- Write the annual mission, quarterly OKRs and active experiments into CLAUDE.md and the files it reads.
- Hold a one-hour weekly retro that checks your spoken notes against Stripe or PostHog numbers.
- Use an experiment template with baseline, window, confounders and kill criteria, and log every result.
- Connect analytics and CRM tools through APIs or MCP so the agent pulls baselines itself.
- Send big plans to a subagent panel with a “what is missing” critic, run on a lighter model.
- Hand agents tasks of one to two hours, and back their code with an AI reviewer and error tracking.