Multi-agent systems
Language models for judgment.
Deterministic code for everything else.
Zeus is an orchestration system for teams of AI agents. Claude handles the reasoning. Routing, safety checks, execution and record-keeping run as ordinary, testable software, so the system can always show what it actually did.
HOW A REQUEST MOVES THROUGH ZEUS
The problem
Agents that sound right are not the same as agents that are right.
Many agent frameworks ask a language model to do everything: remember state, decide which tool to use, enforce its own limits and report on its own work. The result is often confident output that doesn't match what really happened, along with large prompts that are expensive to run.
Our approach
Give the model a smaller, clearer job.
In Zeus, the model is asked to reason, plan and write, which is where it's strongest. Anything with a known correct answer, like an agent's status, whether a task finished or whether an action is allowed, is answered by code that reads the real state of the system.
This keeps prompts short, makes behaviour easier to test, and means a report from Zeus reflects what actually ran.
How it works
A clear split between thinking and doing.
Specialist agents
Work is split across focused agents for research, engineering, browser tasks and quantitative analysis, each with its own tools and permissions.
Targeted context
Each request carries only the information relevant to it, instead of the full system state, which keeps token use low and answers focused.
Rule-based checks
Before anything touches the outside world, it's checked against fixed limits. If a check can't be completed, the action doesn't run.
A record of everything
Task states and results are stored in a persistent ledger, so status reports come from data rather than from the model's recollection.
Platform
Where Zeus is being applied.
We're developing Zeus against real workloads, starting with areas where mistakes are expensive and every action needs to be explainable.
QUANTITATIVE RESEARCH
Market analysis with strict risk controls
Agents analyse market structure and propose trades. Code then decides whether each trade is allowed, and how large it can be.
- Fixed maximum risk per position
- Trading costs checked against volatility
- Trades only in the direction of the larger trend
- Automatic stop adjustment once in profit
RESEARCH & SYNTHESIS
Multi-source research
Research agents gather and compare sources, then return the underlying findings alongside a written summary rather than the summary on its own.
- Sources kept with every finding
- Full results passed back, not placeholders
- Failures reported as failures
TASK AUTOMATION
Browser and workflow tasks
A dedicated task engine completes multi-step web workflows, with each step observed, checked and logged.
- Step-by-step verification
- Loop and stall detection
- Human hand-off when it's unsure
Principles
What we hold ourselves to.
Verifiable over plausible
If the system says something happened, there should be a record showing it did. We'd rather report an honest failure than a convincing guess.
Fail closed
When a safety check is missing or unclear, the default is to stop. This matters most for actions involving money, credentials or external systems.
Precise context
Smaller, relevant prompts are cheaper, faster and easier to reason about than giving the model everything and hoping it picks out what matters.
Built on Claude
Why Claude sits at the centre of Zeus.
Zeus depends on a model that can plan across several steps, use tools reliably, and stay within the constraints it's given. Claude does the planning, delegation and final write-up in Zeus. The surrounding code makes sure those plans are carried out safely and recorded accurately.
As we move from development to sustained use, most of our compute goes to Claude calls for planning, tool use and synthesis across the agent fleet.
Contact
Interested in what we're building?
We're happy to talk with researchers, engineers and potential early users.
founder@zeuscognitive.site