A note from the founder
A model is not a marketing team.
Most AI marketing software begins and ends with access to a foundation model. We started with a different question: what would it take to build a world-class marketing team in software, end to end?
We spent the last year building the missing infrastructure: shared context for humans and agents, a custom agent harness, specialist systems for the major growth functions, and deterministic controls for work that has to be exact. That infrastructure is SynthWeave, the world model our whole company runs on, and Sutton, the AI marketing agents that work from it.
So this is not a model wrapper. Access to a model, on its own, buys you very little.
Models will change. The context, tooling, evaluation and operating judgment around them are what compound. That is the bet we are building, and it is the system we run our own company on.
Brian Bigelow, CEO
01 / The system around the model
More than a prompt. More than one agent.
The context, harness and checks that let a small senior team and its agents run real work.
Persistent business context
The business stays attached to the work. People and AI agents read from and write to one shared place, so what one person or agent learns is there for everyone.
A purpose-built agent harness
It coordinates models, tools and context. Agents work over a live surface of 70+ tools plus an agent registry of 100+ tools.
Checks before work moves
A consequential agent change is held as a proposal by a server-enforced approval gate until a person accepts it.
02 / How the work is built
AI where judgment helps. Code where precision matters.
Interpret the situation
Read signals, form a judgment and adapt the work to the business. The calls stay with people, who answer for them.
- Signal interpretation
- Business-specific reasoning
Enforce what must be exact
Run repeatable checks, protect boundaries and keep a reliable record. Work moves inside the limits a client sets, and anything beyond a limit waits for the client.
- Deterministic checks
- Operating record
03 / The people behind the system
Built by AI engineers and operators.
The people who built SynthWeave come from the places where agents either work or fail.
LlamaIndex
Building the infrastructure agents use to retrieve context, reason and act.
Anthropic
Evaluating model behavior and reliability in production AI systems.
Nike
Operating inside one of the world's most exacting consumer businesses.
04 / For investors and technical teams
Go deeper in a technical briefing.
Investors and technical teams can ask for a briefing on the world model, the agent harness, the isolated sandbox and per-org row-level security.
Where agents run
A runtime built for agents that read, write, and act on a shared record.
SynthWeave is a hosted, multi-tenant platform, live at app.synthweave.ai, with auto-provisioned organizations. Inside it, agents work over a live MCP surface of 70+ tools, plus an agent registry of over 100. They read and write the shared context, they route their own output for human sign-off and poll the verdict, and they run real code in an isolated sandbox. Off-the-shelf agent frameworks give you the loop. What they do not give you is a company-wide record that people and agents both write to, with a gate the platform enforces before a write lands. Pairing those two is why the runtime is our own, and the rest of this page is what that pairing buys.
The record underneath
A structured model of one business that people and agents both work from.
Underneath the harness is the context spine: a company-wide context substrate of snips, bases, and projects with hybrid search, read and written by humans and agents both. It holds what the company knows, what it has decided, and what it is doing, as one permissioned record rather than five copies in five repos. It is the thing the current vocabulary calls a world model, scoped to one business rather than to the open web, and it is a place a person opens, not only a thing an agent queries. Because everyone and every agent works from the same record, the work compounds: what one agent figures out, the next one starts from, and nothing re-onboards at the next stack change. That is the asset, and it is the reason the record is worth more in month twelve than in month one.
- read + write by people and agents both
- hybrid search across snips, bases, projects
One org per customer
Every customer gets a dedicated org. Every Sutton client does too.
The platform is multi-tenant, and every customer gets its own dedicated organization, provisioned automatically. Your context spine, your operating record, and everything your agents produce live in that org. Data is isolated per organization by Postgres row-level security at the data plane. Sutton, our AI marketing agents, is one more customer of the platform, so every Sutton engagement gets its own SynthWeave org the same way: dedicated, provisioned at the start, and the client's from day one.
Held before it lands
When an agent goes to change something that matters, the platform holds it.
A consequential agent write is held as a proposal by a pre-application, server-enforced approval gate, at three independent enforcement layers. The human reviews a proposal, and the write never lands until a person accepts it. Stopped before it happens, not logged after. Enforcement is at the data layer, not written into a prompt, so it does not depend on an agent choosing to honor a rule or on someone remembering to place a node in a flow. Agents can also route their own output for human sign-off and poll the verdict. And every run is a record you can read: what the agent touched, where the halt landed, and what a human accepted. That trace is how you see the blast radius before you let anything act.
Enforcement is server-side at the data plane, not a rule written into a prompt.
A company asset, not a vendor's
It sits above any single model or harness.
SynthWeave is not a wrapper around one vendor's model with a dashboard on top. Your team can run Claude Code, Cursor, Codex, or your own harness, at the same time, and the record stays the one thing they all work from. The sharpest argument against every tool in this category is that the model vendors will absorb it. The honest answer is that a team on multiple harnesses needs a substrate above any single vendor's model: vendor-native memory optimizes that vendor's lock-in, not your company's portability. So the world model is a company asset, and it stays one. The test you should run on any vendor here, us included, is the reversible one: if you deleted the account tomorrow, could your team still ship a change by Friday?
Read the code first
Every run is a record you can read.
The proof is not a claim about reliability. It is the run trace itself. Eight steps from open file to close run, and the one consequential write marked held, waiting for a human to accept. You can read exactly where the halt lands, before you commit.
- snips· the unit of record
- bases / projects· how they organize
- row-level security· per-org, at the data plane
- isolated sandbox· real code runs in E2B
We run our own operation on this. Roughly 50 production tools and live brands, scars included. That is the proof line, and it is ours, not a customer's.