Private design-partner preview

Sooner or later, an agent does something you can’t undo.

It keeps happening to teams like yours: a production database gone in nine seconds. A destructive command run during a code freeze. No confirmation, no warning, no one in the loop. Your agents are getting powerful faster than you can govern them — so Saer is building the layer they act through, where the actions you can’t take back have to stop for a human before they run, not surface in the logs after.

Become a design partner See how it works

Pre-product. We’re building with a small number of founders who’ve already had the scare — and don’t want the next one.
Button didn’t open your mail app? Write to [email protected].

The problem

The more your agents can do, the more your control logic falls apart.

Every new agent gets its own permission checks, prompt rules, approval hooks, and logs. None of it adds up to a system. The more useful your agents get, the more brittle and expensive control becomes.

The old way: control rebuilt per agent

Permission, review, and memory get re-invented every time you add an agent.

  • Permissions, prompt rules, and approval logic scattered across your code.
  • A human babysits anything risky, because nothing is trusted unattended.
  • Logs tell you what happened, after it already happened.

The layer we’re building

One governed operating layer your agents act through, instead of bespoke guardrails.

  • Agents act through shared permission, evidence, and memory.
  • High-stakes actions become proposed state until someone with authority accepts them.
  • Every action carries intent, authority, evidence, and consequence.
Operating model

Not a chatbot.
Not after-the-fact observability.
A governed layer your agents act through.

Saer turns scattered company state into a system agents can reason from, act through, and answer back to — with control native, not bolted on.

01

Operating Graph

Typed work state: people, tools, entities, decisions, sources, permissions, actions, and consequences.

02

Governed Memory

Bounded, provenance-rich context for agents. Memory with control, not loose summaries.

03

Command Layer

Human intent becomes governed action across agents, tools, and time, under permission and review.

04

Evidence & Review

Findings, approvals, traces, audit records, confidence, contradiction, and rollback boundaries.

Operating loop

From signal to action to memory — without losing the reason why.

SenseIngest signals from tools, docs, meetings, markets, and systems.
ThinkSurface options, contradictions, risk, uncertainty, and evidence.
DecideCreate findings and decision records linked to source material.
ActDelegate bounded tasks through agents, workers, and tools.
ReviewRoute high-stakes mutations through approval and audit gates.
LearnFeed consequences back into memory, evals, and future workflows.

Observability tells you what an agent did. Saer governs what it’s allowed to do — and why.

Govern, don’t just observe
Permission-aware contextAgents should only see and use the state they are allowed to act on.
Human review gatesHigh-stakes work becomes proposed state until accepted by the right authority.
Evidence trailsDecisions carry source, freshness, confidence, contradiction, and limitation metadata.
Governed memoryMemory mutation is explicit, reviewable, reversible, and bound to consequence.
Action recordsEvery side effect can preserve intent, actor, permission, tool, result, and outcome.
Model/tool neutralityThe OS governs work across tools and models rather than depending on one assistant.
Who this is for

For founders who’ve already had the scare — or can feel it coming.

This is you if…

  • An agent in your company already touches prod, money, data, or customers.
  • You rebuild the guardrails by hand every time you add one.
  • You’ve read the nine-second horror stories and thought: that could be us.
  • You can feel capability getting ahead of control.
Private preview

Help shape the governed operating layer for agent-heavy companies.

Saer is pre-product and building with a small number of founders already living this problem. If your agents are getting useful faster than you can govern them, let’s talk.

  • You give: an hour a month, and the truth about your agent incidents.
  • You get: early access, a direct line to the builder, and a say in what the gate refuses first.
Become a design partner

or write to [email protected]

Who’s building this

Saer is built by Richard Woodman in Cardiff, UK — hands-on with coding agents daily. Related public work: eval-oracle, an eval-gated autonomy harness for AI agents with an audited judge, and more at github.com/wlshlad86.