Glasshat / Evaluation prototype
Watch demo ↗A judgment needs evidence.
Evidence → specialist perspectives → audit → report.
Parallel perspectives, an audit, and a bounded correction step make the reasoning available for review.
Conceptual company structure
- Human direction
- Agent departments
- Business operations
- Accountability
The next workforce is software.
Build a company.Deploy its workforce.
AgentBase builds and operates AI agent fleets for real business operations.
We run client workloads, build agent fleets inside companies, and prove new systems in two-week engagements.
Run · Build · Prove01 / Social Seeding · Creator campaign operationsSocial Seeding ↗
A brief goes in.
A team gets to work.
One campaign brief, seven specialist roles. Social Seeding gives discovery, outreach, replies and reporting a shared context—with people reviewing the decisions.
The brief becomes shared context.
Original V2 interface, rendered from the product source with empty example fields.
View the interface ↗Closed beta. The V2 walkthrough shows agent roles at different stages of implementation; people review outreach drafts before sending.
02 / kbeauty.market · Video-first discoverykbeauty.market ↗
The video is
the front door.
Start with a creator or brand clip. Find the product behind it, inspect the details, and keep a path back to the source. Kbeauty connects discovery to context.
A video-first storefront connects clips, their original sources and product information in one place.
The opportunity: change how people discover products, then build the operating tools around that experience.
Storefront browsing preview. Own checkout is not yet enabled.
03 / Memex · Memory for agent workMemex ↗
New session.
Not a blank slate.
Memex makes past agent sessions retrievable. When work gets stuck, Loop Breaker suggests a next action from a similar session—and opens the exact turn behind it.
Repeated build failure → retrieve a similar session → inspect the recorded next action. Interactive illustration using a synthetic Rust CI session.
Memex provides local memory for Claude Code and Codex through MCP and session hooks.
This is a synthetic session walkthrough, not a live retrieval run.
04 / Make the reasoning and the handoff inspectable
Show your work.
Show your work.
Carry it forward.
A useful agent leaves more than an answer. Our tools explore how to examine a judgment, make a choice visible, and carry that decision into the next piece of work.
PreviewForge / Development workflow
Watch demo ↗A choice becomes a brief.
Explore visual directions → record the choice → hand off to the specification.
A Claude Code workflow carries a selected preview and its decision record into specifications, implementation and tests.
Selected work from our team, including projects begun before AgentBase LLC. These products and tools are building blocks for our ambition; they are not yet one autonomous company.
05 / From a responsibility to an organization
The ambition is
The ambition is
an entire organization.
Our goal is to replace human operating labor with software organizations. From one assignment to connected departments, agents take on the work. People set the direction and remain accountable.
Conceptual task structure
- Brief
- Research
- Summary
Stage 01 · Task
Complete a defined assignment.
Read the brief, use the relevant context, and return a result someone can inspect. Define the input, the allowed work, and what counts as done.
In practice, for example
Illustrative assignment: research a creator against a campaign brief and return the evidence for their fit.
06 / The capabilities behind the work
Capabilities
Capabilities
we build with.
The building blocks behind the work: specialist roles, persistent memory, inspectable judgment, and execution.
01Coordinate specialist agentsDefined roles, shared context, durable handoffs.Role tools · workflow checkpoints
02Draft from supplied factsStructured outputs that stay connected to the brief.OpenAI · tool calling · structured outputs
03Retrieve useful experienceFind a past action and inspect its original context.Memex · Qdrant · multi-vector retrieval
04Carry context across toolsBring source-linked session memory into the next task.MCP · source-linked memory
05Inspect and audit judgmentsKeep the evidence and correction process visible.Google ADK · parallel evaluation · bounded correction
06Turn decisions into specificationsMove from a visual choice to a reviewable specification.PreviewForge · decision records · artifact checks
Current fleet models
Google Gemini OpenAI
Capabilities demonstrated across our products and tools.
07 / Work with AgentBase
Bring us a job
Bring us a job
that needs doing.
Start with a recurring operation, a process to change, or a system to build. Choose a working model below.
01 / An operation to run
Run
We operate recurring workloads with an agent fleet and clear decision points.
02 / A company to equip
Build
We embed an agent fleet in your company, with defined responsibilities and human oversight.
03 / A system to prove
Prove
We turn a bounded problem into a working system in a two-week engagement.
Start with an example input, the decisions a person makes, and the output that counts as done.
contact@agentba.se
Email us — both founders read every line.
Meet Theo, the outreach writer. Campaign facts become an email draft; a person decides whether it can be sent.