Now, next, and not yet.

• The guided action loop: npx aibill improve finds a waste pattern in your own local evidence, drafts one reversible test, records your typed APPROVE before anything changes, and calculates one canonical matched-session, quality-gated result. improve --sample is the labeled practice run; it writes nothing.
• The agent-native loop: the read-only draft_improve_command MCP tool lets an AI client draft the change/rollback/canary plan conversationally and hand over one paste-safe improve --draft command. Approval is still typed by the human in the terminal, never by an agent.
• Local accountability contracts: identify records explicitly confirmed ownership, outcome github opt-in verifies a merged PR whose observed checks passed, and accountability answers owner, outcome, approval, and measured result from one private view. These are foundations for Workspace, not company identity or approval routing.
• Claude Code and Codex local evidence in the CLI, explicit MCP/plugin, private init/cache, and optional cache-only Claude Code statusline.
• Separate provider-billed cost, subscription context, API-equivalent value, validation coverage, freshness, and missing coverage.
• Optional OpenAI and Anthropic provider reports; Cursor and GitHub Copilot connectors remain fixture-verified beta.
• An experimental, fixture-verified Gemini CLI financial reader. It is financial-only and never enters statusline, Glance, Context Health, Apply, plan, runway, or invocation evidence.
• An additive Agent Economics Receipt v0 contract in @agent-finops/core, plus a registry and generated-doc foundation for local readers.
• An unsigned, source-built Glance preview. The public Mac download and shared Workspace are not launched.

logs.json is detection-only and creates no financial row. Gemini does not feed statusline, Glance, Context Health, plan, runway, invocation evidence, recommendations, or Apply.

1. 1

   Deepen attribution and session evidence before outcomes

   Add branch, ticket, and work-unit attribution plus plan presets and session vitals before accepted/rejected/reverted/rework outcome states. The first verified unit is an accepted coding task or merged PR—not lines of code or modeled hours.

2. 2

   Add financial CI without autonomous enforcement

   Start with warnings, preview, dry-run, explicit approval, rollback, and verified result. No autonomous control claim without a real adapter and validated team policy.

3. 3

   Close distribution and comprehension proof

   Run the 8–12-person study, collect real billing-reconciliation cases, and complete signed, notarized, safely updated Glance distribution before offering a public download.

4. 4

   Earn the organization foundation with design partners

   Test one read-only observability import and gather design-partner proof first. Then begin an opt-in aggregate receipt sync for shared reconciliation, allocation, and approval history. This is a foundation—not Workspace general availability.

The intended company product is an opt-in, permissioned Workspace over approved Agent Economics Receipts: organization-wide reconciliation, project/client/user/agent allocation, budgets, anomalies, approvals, audit history, accepted-outcome economics, and a read-only financial teammate that cites its evidence and missing coverage.

Defensible ROI needs reconciled cost, an accepted outcome, and independently evidenced business value. Token volume, activity, lines of code, or invented hours saved are not ROI.

• ROI or productivity claims from usage evidence alone.
• Autonomous budget enforcement or provider changes.
• Cursor-local financial parsing from current internal IDE databases; official admin APIs remain the financial path unless a stable, versioned local format emerges.
• General Workspace availability, RBAC, or team billing.
• A signed public Glance binary.
• A general trace explorer, prompt warehouse, model gateway, or long-tail parser race.