A fleet, not a chatbot

Supernova models agents as an operating workforce. An agent is not just a prompt attached to a model call: it has an identity, role, capability set, session history, diagnostic surface, project scope, and a place in a durable workflow.

Provider-neutral by design

The runtime uses one canonical representation for turns, tools, and results. Provider adapters normalize the transport details. Adding a model should mean adding an adapter—not rewriting the agent loop, permission system, or tools.

The intended production runtime is a thin, owned Rust loop rather than a wrapper around a vendor’s interactive coding application. Generic coding tools, Supernova’s typed domain tools, and MCP tools all enter through the same observed registry. Every tool call can therefore be attributed, checked, and emitted into the platform trail.

Token economics is a runtime concern

The limiting resource is not only dollars. Subscription allowance, uncached input, agent turns, context reloads, latency, and follow-on work all affect how much useful work the fleet can deliver. Supernova therefore treats token economics as architecture:

• warm-session forks are the default; cold boot requires an explicit fallback;
• stable prompt and project context stays at the front of the cached prefix;
• repository updates append as deltas and periodically roll into a new warm base instead of forcing a full context reload;
• easy questions use sun find or the memory graph before spawning research;
• large-context questions fork a preloaded, never-compacting oracle for one answer, then discard it;
• agents attach to events and sleep rather than spending turns polling;
• every goal has iteration, wall-clock, and token ceilings with a structured soft-stop;
• work selection records expected outcome, P50/P90 token cost, and realized value per token.

Monitor owns measurement: per-provider and per-account usage, spend rate, budget pressure, turn receipts, and cache field metrics. Agent owns avoidance: choosing the provider, account, warm parent, lookup path, or model appropriate to the job. This keeps cost dashboards from becoming a second scheduler.

Fork is the basic spawn primitive

A warm parent session can fork child agents with its prompt context already cached. Different jobs— implementation fan-out, reviewers, investigators, or specialist subagents—become policies around the same primitive.

flowchart TD
    parent["Warm parent<br/>project context + cached prompt"]
    parent --> coder["Coder fork<br/>one worktree"]
    parent --> reviewer["Reviewer fork<br/>independent authority"]
    parent --> investigator["Investigator fork<br/>failure context"]
    parent --> oracle["Oracle fork<br/>bounded question"]

This is both an efficiency feature and a coherence feature. Children begin with shared context while receiving different permissions, evidence obligations, and termination conditions. Field experience in the predecessor system demonstrated roughly 99.7% cache reads and about $0.008 per warm-oracle question. Supernova turns that useful hack into an explicit runtime contract; matching production field evidence for the final runtime remains an open admission requirement.

React, do not poll

Agents attach to typed event conditions and sleep at zero model cost. A matching bus event wakes the right session. The same parked-await idea serves agents and durable workflows, so waiting for review, recovery, approval, or new evidence does not require a polling prompt loop.

Goals terminate on evidence

A goal binds a target, an evidence obligation, and a budget. The loop is deliberately small:

act -> check existing evidence -> park -> wake -> repeat

The check reads the platform’s derived closure result; the agent cannot declare its own work done or manufacture a private definition of success. Iteration, wall-clock, and token ceilings make failure loud and return a structured result to the parent.

Each pass starts from the evidence produced by the previous one. The agent revises its next action against the remaining delta and recursively repeats until the specified goal state is demonstrably satisfied. Parking makes that recursion durable and cheap; it is not an unbounded prompt loop.

Diagnostics instead of watchdog kills

Each session exposes what it is doing, its recent activity, runtime placement, health, tool use, and budget state. The operating rule is diagnose first. A slow turn is not killed merely because a timer expired; termination is an explicit lifecycle action with a recorded reason.

Incentives are attached to responsibility

Agent roles are separated so that improving a personal score cannot mean weakening the bar. A reviewer does not implement or approve the work it reviews; an outcome steward cannot close work by narrowing the outcome; and an implementer cannot change the evidence obligation after readiness.

The planned reviewer scorecard is a visible vector of responsiveness, comprehensiveness, avoidable round trips, independently attributed escaped defects, false-positive burden, calibration, and process integrity. Raw speed, comment volume, and finding count carry no positive weight by themselves. Results are normalized against comparable work and always retain sample size, uncertainty, evidence, and policy version.

Activation moves from measurement-only shadow mode to advisory use and only then to constrained routing. Adequately supported outcomes may affect the classes of work an agent is trusted to review, but a scalar score can never grant a capability or bypass an eligibility floor. Monitor audits bias, metric gaming, risk avoidance, fragmentation, and attribution disputes; it may propose a policy change, but cannot apply one.

Memory with boundaries

Per-agent memory and derived knowledge indexes live in the agent system. Project truth does not. Agents can rebuild or discard their own convenience state without losing requirements, issues, evidence, or code. That separation prevents an agent’s private recollection from becoming a hidden source of authority.

What is working now

Supernova already runs project-scoped agent sessions, records lifecycle and tool evidence, supports goal and wakeup surfaces, and uses its agent machinery during its own construction. The final owned multi-provider runtime and some provider-network paths remain incomplete; current operation includes interim adapters. See project state.