Mixture of agents

Chronos is agent- and model-agnostic. Each session runs one coding agent in its own isolated worktree, and you choose which one — so different tasks, teams and preferences can use different agents side by side in the same workspace.

Choosing an agent per session

When you create a session, pick the agent that runs it:

• Claude Code
• Codex
• OpenCode

Every agent runs behind the same adapter interface, so sessions look and behave the same regardless of which agent is inside: the same stream, the same diffs, the same audit trail, the same governed context and connections. Adding support for a new agent is an adapter, not a platform change.

Model and reasoning effort

Alongside the agent, each session selects:

• Model — which model the agent uses for the session.
• Reasoning effort — how hard the model thinks before acting. The selector maps to each agent’s native control (for example Claude Code’s effort setting or Codex’s reasoning-effort configuration), so you tune depth versus speed per task: low effort for routine edits, high effort for gnarly debugging or design work.

Both choices are per session — running one session on a fast model and a parallel one on a deeper model is the normal way to work.

The direction: governed multi-model aggregation

Beyond one-agent-per-session, Chronos is heading toward mixture of agents: fan a turn out to several reference models that each contribute analysis, while a single aggregator model reads their opinions, writes the actual response and makes the tool calls. References only advise — exactly one model acts, so there is never a conflict of parallel edits to reconcile.

What makes this Chronos rather than a raw fan-out is governance:

• Governed presets. Which reference and aggregator models a preset may use comes down signed from the enterprise panel — a user cannot point a turn at an arbitrary provider, because every reference model is another party seeing your code.
• Budget gates. Aggregation multiplies token cost several times per turn, so presets run under per-session and per-team budgets with approval above threshold.
• Full audit. Every reference and aggregator call is recorded with cost attribution — never silent.

The sweet spot is read-mostly reasoning where a second opinion earns its cost: spec and architecture work, and above all code review and the certifying agent in the merge gate. It is deliberately not a default for routine editing turns.

This composes with a pattern Chronos already supports natively: best-of-N parallel sessions — run several independent sessions on the same task, each in its own worktree and branch, and pick or merge the best result through the merge gate. Aggregation parallelizes opinions within a turn; parallel sessions parallelize whole attempts. Together: N candidate implementations, scored by a multi-model reviewer.