A note on cost, first
ZO v1 is optimised for production-grade research engineering. Agents iterate, verify, and write production-quality code, and that burns tokens. A canonical reference run with the default profile is on the order of $10–15 of API credits. For cost-sensitive plans (Anthropic Pro, student tier, individual researchers on a budget), the--low-token preset currently delivers ~30% measured savings on the same workload. We’re actively pushing that further, details below.
The v2 rearchitecture
The headline: after a full-source review of three agent-orchestration systems — oh-my-claudecode, ruflo, and ralph — ZO v2 adopts 12 features across five workstreams. ZO’s oracle discipline stays; underneath it we’re adding:An enforcement plane below the prompts
Deterministic hooks that mechanically guarantee what spawn prompts currently promise: deliverable contracts checked when an agent stops, memory that persists automatically, gate approvals that can’t be forged, verifiers that structurally cannot edit code.
A machine-readable control plane
A plan ledger where every subtask has verifiable acceptance criteria and a pass flag only the oracle may flip — progress becomes a query, not prose interpretation. STATE.md stays as the human-readable projection.
A fresh-context execution substrate
A new agent per iteration, all state on disk, git as the checkpoint — context rot and compounding errors eliminated structurally, with a watchdog built from battle-tested parts (heartbeats, never-fight-these-stops taxonomy, rate-limit auto-resume).
An oracle for the platform itself
Every learned rule binds to a code marker checked in CI; every documented failure gets a smoke test reproducing its symptom. The self-evolution loop gets the same rigor ZO applies to ML work.
What we’re working on
Alongside the v2 rearchitecture, three ongoing tracks shape the next 1–2 quarters, driven by recent user feedback on cost, accessibility, and cross-phase autonomy — they fold into or queue behind the v2 workstreams above. Grouped by horizon: Now, Next quarter, and Major design.Now (next 2–4 weeks)
Caveman integration
Bundle the caveman terse-output skill into
--low-token mode, scoped to agents that don’t emit structured artifacts (data work, summarisation; not the oracle or training-callback paths). Target: another 10–20pp on top of the measured ~30% savings. First validation of our planned-integrations promise.Onboarding hardening
One-line installer (
curl … | bash), Homebrew formula, and a new zo doctor command for auto-detect-and-fix on missing dependencies. Reduces the setup-friction wall before we invest in a full UI.Next quarter (1–3 months)
Cross-phase backedge primitive, DAG to graph, step one
Today’s orchestrator behaves like a directed acyclic graph: phases run forward, gate-by-gate. When an issue surfaces during modelling that traces back to preprocessing, for example, a flipped waveform polarity that’s invisible in raw data but obvious in trained-model behaviour, it requires human steering to revisit earlier phases.What ships: a new
RETURN_TO_PHASE gate decision lets a later-phase agent autonomously request a revisit of any earlier phase, with budget guardrails (--max-cross-phase-revisits N), dead-end detection, and a revisit-reason payload that re-prompts the earlier phase with the new constraint. Validates the weighted-graph hypothesis cheaply, before the full refactor.VS Code extension, agent dashboard, comms feed, gate approval
The CLI-first experience suits software engineers. Researchers from other disciplines often hit friction with environment variables, Node.js, and tmux. The answer: a VS Code extension (not a fork) that gives ZO a visual home alongside the editor, file tree, git, and integrated terminal researchers already have.Architecture: headless tmux as the engine, xterm.js webview as the renderer. Claude Code keeps its real TTY (zero regression risk on the orchestration core); the UI gets agent status panels, a comms feed, training-metrics dashboards, and one-click gate approval, all in one window with the user’s code. A “ZO Studio” bundled installer (VS Code + extension + Claude CLI + uv, one-click per platform) is on the table as a follow-on for non-technical researchers.
Push --low-token toward 70–80% savings
Current preset gives ~30% measured savings (see cost benchmark). The path higher is architectural: prompt caching via direct Anthropic SDK integration (replacing the
claude CLI subprocess), Batch API for non-real-time phases, Files API for shared context across the team. Multi-week effort, gated behind the SDK refactor.Major design (3–6+ months)
Full weighted graph architecture
Phases as nodes with weighted reversible edges. Each node carries a “completeness” probability that updates as work progresses; the system ranks revisit weights and runs a prioritised ablation queue. Users can freeze any node manually if happy.Enables a fast, loose end-to-end first pass, then the system autonomously revisits earlier nodes with better project context, instead of requiring human steering. Gated behind the cross-phase backedge primitive landing first.
Broader coding-agent provider support
Today ZO orchestrates Claude Code. Adding adapters for other coding-agent runtimes, OpenAI Codex CLI, OpenCode, others, would let users bring their preferred provider and hedge single-vendor risk.Realistic scope: worker-pool model (Claude Code lead dispatches single-agent jobs to alternate workers), not full peer-to-peer team parity. Full parity would mean rebuilding Claude Code’s native
TeamCreate / Agent / SendMessage infrastructure on top of bare APIs.Where to weigh in
These priorities come from real user feedback. If you have a use case one of these would unlock, or a different priority you’d argue for, open an issue or join the discussion.Issues
Bug reports, feature requests, prioritisation discussion.
Discussions
Use-cases, design questions, “would ZO work for…” threads.