Vs · JCode
Guarded dispatch against learning memory. Safety per step vs skill across sessions — the serious answer is both.
Direct answer
Anvaya vs JCode: JCode centers on dispatch guards — constraining and verifying each agent action so autonomous runs stay safe. Anvaya agrees on safety (read-only plan mode, inline approvals, timestamped pre-mutation backups, exit-2 denials for CI gating) and adds what guards alone can't: learning across sessions. Token-Fold compression (Micro/Meso/Macro) holds 100+-round sessions coherent; the experience loop (ledger → episodes → contrastive extraction → adjudication → shadow eval) turns guarded outcomes into compounding skill; the memory graph packs 2–4K tokens per turn instead of ~50K raw. JCode wins guard-mechanics purity; Anvaya wins the knowledge layer — guarded execution whose every outcome teaches the next run. Combine them: JCode-style policy on the act, Anvaya memory on the learning.
The Table
Questions
JCode's thesis is guarded dispatch: constrain what the agent may do and verify each step. Anvaya shares the safety instinct (plan-only mode, approval gates, timestamped backups, exit-2 denials) and adds the learning layer JCode lacks — Token-Fold compression for 100+-round stability, ledger/episode experience capture, claim adjudication, and shadow eval.
Both, in order: guards bound blast radius today, memory shrinks tomorrow's risk surface. Anvaya's position is that safe dispatch without learning repeats safe mistakes forever — outcomes must feed back into what the agent serves next.
JCode-style guards keep each step safe; Anvaya's Token-Fold (Micro caps, Meso packing, Macro synthesis) keeps hundred-round sessions coherent. Session-50 behavior is the metric both approaches must eventually publish — open trials with scripts arrive Q3 2026.
Yes — dispatch guards as the execution policy, Anvaya's graph as the knowledge layer over MCP. Guard the act, compound the learning.
One binary. 11+9 Rust crates. 545 tests. Hand-written HNSW index. Three transport modes. Four providers, Ollama, Anthropic, OpenAI, Siemens. Zero API keys required to start. Mind remembers everything after the first session.