How it is built
One shared record, a gate the platform enforces, and an honest account of the rest.
The problem SynthWeave exists for is not that agents forget. Storage is solved, and you solved it. It is that the moment more than one person, more than one agent, or more than one harness depends on the same context, nobody is working from the same thing, and the rules you wrote turn out not to bind. This page is how the platform answers that, in full, including where it stops. If a claim here has a limit, you will find the limit next to the claim, because that is the only version worth reading.
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 self-serve signup and 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.
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. The honest shape of that, stated before you ask: all organizations share one process and one set of workers. SynthWeave is org-scoped, not org-isolated. The only per-organization isolation is at the data plane, through row-level security; there is no isolation at the level of compute, memory, or filesystem. Sutton, our forward-deployed marketing offering, 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.
One limit worth knowing up front. The automatic gate covers edits to existing documents (snip_rewrite and snip_patch). Creating new documents and changing metadata are not gated yet, and it depends on auto-accept being off.
Where we stop, on purpose
What the platform does not do yet, said plainly.
The comparison table on the homepage has a real loss column for a reason. Here is the same honesty about the platform itself. None of this is softened, and where an objection lands on the frontier, the honest answer is that it is currently unanswerable, and we will say so rather than invent a rebuttal.
- Org-scoped, not org-isolated. One process, shared workers. The data plane is separated by per-org row-level security; compute, memory, and filesystem isolation between orgs is not something we claim.
- No SOC 2, no ISO, no certification, and no trust center. If your review eliminates on a questionnaire before the conversation, we are not the tool yet, and we are not apologizing for the absence of a certificate.
- No self-hosting, and no data-residency guarantee.
- No app store, and no third-party app execution. Real code runs in an isolated sandbox.
- The workflow engine runs an ordered list of steps. No branching, no looping.
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
Do not take the architecture on faith. Read a run.
The proof of all of this is not a claim about reliability. It is the run trace itself: eight steps from open file to close run, with the one consequential write marked held, waiting for a human to accept. You can read exactly where the halt lands before you sign up, and you can try to make an agent overwrite something and watch whether the write lands. If it lands, we did not do our job.