Follow along: I share what breaks (and what fixes it) putting AI agents into production — @surlymochan on X · Contra

⭐ If evalid or swrm saves you time, a star is the best way to support the work — I'm pushing for 100 stars this month.

AI agents · Harness engineering · Reliable loops & graphs

I build tools for running coding agents, coordinating their work, and checking what they deliver. My focus is executable workflows, isolated execution, and experiments others can reproduce.

All pull requests I opened or co-authored — in my own repositories, in nibiru-labs (where my daily work lands under my nibiru-anu alt with me as co-author) and in external projects: total opened (light blue) vs. merged (dark blue), bucketed by Singapore date. Charts refresh hourly.

• 🌐 Phanes — my current flagship: a distributed, heterogeneous, self-evolving agent network. One control plane for goals and acceptance, a unified pool of compute resources, and final acceptance that must cite independent verification — never a worker's self-report. Architecture passed an adversarial multi-model design review; building in stages.
• ✅ evalid — the runtime verification layer for coding agents: when an agent claims its tests pass, evalid replays the claim against a sealed full-tree snapshot — claimed ≠ verified. npm
• 🐝 swrm — a distributed dispatcher for heterogeneous agent tasks: workers claim tasks under leases and return evidence, across agents and model families, each task in its own git worktree. Getting started

• Phanes is where OpenBot, swrm (its task dispatcher), and my RSI (recursive self-improvement) work converge into one system: a single control plane holding agent identity, goals, authorization, and final accept/reject decisions; distributed heterogeneous execution scheduled against a unified resource registry (hosts, backends, models, quotas described separately from agents); and an evidence-driven improvement loop that proposes changes from verified run records. Tasks flow through one versioned contract — Goal → Task → Run/Attempt → Evidence — with idempotent dispatch, so a lost receipt or a restarted worker can never silently run the same attempt twice.

• evalid is the runtime verification layer for coding agents. It seals the green test run as a full-tree snapshot (per-file hashes plus a run binding), then replays it from the snapshot — a pass only counts if it reproduces, and test files must hash-match the sealed run. Underneath sits a zero-judge pipeline (SPEC clarification, code-review, and test gates) for developers who want evidence, not self-reports.

• swrm (formerly OpenSwarm) is the distributed dispatcher for heterogeneous agent work: tasks go out under claim/lease semantics and come back with evidence, routed across Codex, OpenCode, DeepSeek, and opt-in backends. Each task runs in its own worktree with explicit limits.

Recent merged contributions:

• senpi — weighted CJK text in compaction cut-point budgets.
• oh-my-openagent — named the parked-facts cause and run log in /doctor and reported retry outcomes.
• oh-my-openagent — settled the relaunched cache before asserting on the store dir in cache-persist tests.
• opik — accepted repeated identical JSON objects in the TS LLM-judge parser.

Currently contributing fixes to oh-my-openagent, OpenBot, and strands-agents/harness-sdk. Open pull requests are proposals, not merged features or maintainer roles.

I publish benchmark methods and run records alongside my tools. Results are specific to the tasks, models, and configurations tested; small samples need careful interpretation.

• OpenFlow benchmarks
• OpenSwarm experiments
• Merged upstream PRs

More tools will appear here as their installation, documentation, and examples are ready for others to use.