Zeus Features — Safe, Verifiable AI Coding

What makes Zeus different

Six capabilities. Six different failure modes.

Each one addresses a practical limit of agentic work: unsafe changes, unsupported targets, closed integrations, unavailable agents, single-model judgment, or budget-blind execution.

01Containment

Write work stays inside the sandbox.

Each write-capable agent gets a dedicated Git worktree. If Zeus cannot prove a delegate's filesystem boundary, that delegate cannot implement.

02Native testing

Zeus can test what it builds for iOS.

Zeus runs xcodebuild test on a generated disposable Simulator, deletes it after the run, and retains the .xcresult evidence for review.

03Extensibility

Extend the system, not just the prompt.

Add OpenAI-compatible endpoints today. Delegate packs and platform toolpacks provide defined extension points for new CLIs, project types, and checks.

04Resilience

Manual councils keep every seat you choose.

Auto mode selects up to three design reviewers; Manual mode preserves every selected eligible member. When a provider fails or hits quota, Zeus can rotate across the live pool.

Availability depends on the providers, models, and capacity you connect.

05Judgment

More than one model judges the result.

Zeus combines reviewer scores, confidence, blockers, and uncertainty into a risk-aware posterior. It models correlated reviewers and treats disagreement as a reason to seek more evidence.

06Budget guidance

Budget guides the run. It does not kill useful work.

Zeus forecasts a phase envelope and meters live spend without terminating a delegate that is still converging. It prices cached reads, cache writes, uncached input, and output separately.

Built for council workloads where repeated context can make cache reuse a large share of input.

Everything stays inspectable

Seven views answer the questions that matter.

01Goal & scope
02Design review
03Model roster
04Council & dissent
05Budget & stops
06Tests & evidence
07Review & apply