An older, white-bearded Poseidon stands facing us in a stormy sea, resting against his staff with both hands. He wears a simple belted toga and a crown seated on his head with its front band concealing his eyes; fine blue grains form the scene.

THALASSA / BY KORALLI

IN DEVELOPMENT

Poseidon.

Coordination for agent systems.

SEE HOW IT FITS TOGETHER ↓

AI for critical business processes.

Thalassa brings specialist agents, model routing and human review into one workspace.

We’re building it for processes where errors cost time, money and trust. Poseidon coordinates the work, with the aim of making agent systems more reliable and efficient.

EXPLORE THE PROTOTYPE

Interactive workspace and worked examples.
No model execution in the public demo.

SEE HOW IT FITS TOGETHER ↓

A clear route through the work.

Give each agent a clear job. Choose a suitable model for each task, and keep the evidence available for people to check.

01 / ASSIGNMENT
Review a proposed changeChange request · service records
PoseidonAssign responsibilities
InvestigateRead the evidence
AnalyseExamine the work
ReviewChallenge findings
SPECIALIST ROLES
Clear responsibilities, shared context.
02 / MODEL ROUTING
ResearchAgent task
AnalysisAgent task
ReviewAgent task
Model routingFit · evaluation · cost · time
FastShort, bounded tasks
ReasoningDeeper investigation
SpecialistDomain requirements
INTERCHANGEABLE PROFILES
A role can use different models as the task changes.
03 / REVIEW
Source evidenceRecords supplied for the task
ChecksCitations · gaps · conflicts
Human reviewAccept or request changes
Reviewed resultFindings for a decision
If unclear,
gather more
evidence
Uncertain findings return for more evidence.

Illustrative workflow · proposed architecture

THE WORKSPACE

See what sits behind a result.

OPEN THE WORKED EXAMPLE ↗
Thalassa with Poseidon’s team allocation, task and compute controls, and an inspector for the proposed route.
INTERACTIVE INTERFACE PROTOTYPEExample material. No live company systems connected.

WHAT WE ARE TESTING

Reliability has to be measured.

We want to make missing evidence easier to spot and straightforward tasks less costly to run. To test that, we will compare missed issues, unsupported claims, review time and cost against a single model doing the same work.

The prototype shows the intended workflow. Model routing and learning from outcomes remain to be built and evaluated.

DISCUSS A PROCESS WITH US ↗