Build your self-driving GTM engine
The session where Cargo launched programmable GTM infrastructure: the five levels of autonomy, why your coding agent forgets everything overnight, and a repo built live that it can read tomorrow.
An audience already running Claude Code, Codex and Cursor against their go-to-market, asking the same question in six different ways: where does the work go, so that the next session can read it?
- 3:47Go-to-market is splitting into two tracks
The familiar one builds the engine and then runs it. The second hands agents the problem space with the context attached and lets the loop work out which motion to build. Tariq Minhas's argument is that the two converge, and that the second is where the AI-native companies already are.
- 5:06A coding agent with no runtime starts from zero every morning
Your agent calls a CLI, calls an MCP server, gets a result and forgets it. Close the session and the next one re-pulls everything before it can answer a question, which is why people keep one session alive until it hits the context limit. Nothing about the work is retained anywhere it can be read again.
- 9:03Five levels, and most teams are at L2
Borrowed from self-driving. L1 is clicking through a UI. L2 is a coding agent implementing what you describe, which is where most of the room was. L3 is a declarative repo the agent reads and changes. L4 is the agent proposing the work itself as pull requests. L5 is those merging against goals you set, with nobody reading the diff.
- 17:05Four folders are the whole structure
Plan holds the initiatives, context holds the ICP and the personas and what the company knows, cadence is the audit trail of what changed and why, and infrastructure is the workflows and sequences as code. The agent reads files instead of re-deriving the company from scratch.
- 32:48Cargo runs itself at L4, and the jump took longer than expected
Hand an agent autonomy before it has a record of what matters and it proposes work that makes no sense: the wrong accounts, the wrong market. What closed the gap was feeding it everything, internal calls and customer calls and Slack, so its suggestions rest on something. Six or seven weeks in, it takes outbound, customer success, and product pull requests off the logs.
- 36:48Say what the goal is, or it finds the fastest path to it
Told to increase revenue and nothing else, the engine goes looking for the shortest route, which might be calling your customers and asking them to spend more. The constraints in the plan folder are what buy a sensible medium term, and they are also what shortens the agent's learning period.
Two people and a live build: one framing where this is going, one opening a terminal and wiring it up.
The argument and the build. Five levels of GTM autonomy borrowed from self-driving, the four folders an agent reads instead of re-deriving the company every morning, and then a terminal: the CLI, a repo scaffolded live, three cookbooks installed into it, and the Cargo workspace those cookbooks deploy to.
Watch this talkThe framing and the questions. Why go-to-market is splitting into a track that builds the machine and a track that hands agents the problem space, and then the awkward ones from the room: how this lands in a 100-person org, where it goes in six months, and what a team should automate first.
Watch this talk