emerge

A next-generation agent harness — model-agnostic, multi-agent native, and self-aware.

Status

v0.1.0 published to npm 2026-04. Real-model end-to-end verified against a custom OpenAI-compatible gateway (gpt-5.4). Anthropic / direct OpenAI / local-model paths (Ollama, vLLM, llama.cpp) are shipped but not yet verified end-to-end with a real model — see VERIFICATION.md.

Install

Core packages:

npm install @lwrf42/emerge-kernel
npm install @lwrf42/emerge-agents
npm install @lwrf42/emerge-provider-openai-compat   # for any OpenAI-compatible endpoint
npm install @lwrf42/emerge-provider-openai           # for direct OpenAI API
npm install @lwrf42/emerge-provider-anthropic        # for Anthropic Claude
npm install @lwrf42/emerge-sandbox-inproc            # in-process sandbox (fast, no Docker)
npm install @lwrf42/emerge-sandbox-harbor            # Docker sandbox (strong isolation)
npm install @lwrf42/emerge-tools                     # fs + bash tools
npm install @lwrf42/emerge-telemetry-jsonl           # JSONL session logging
npm install @lwrf42/emerge-replay                    # session record/replay
npm install @lwrf42/emerge-surveillance              # adaptive decomposition

Interactive tools (install globally):

npm install -g @lwrf42/emerge-cli @lwrf42/emerge-tui @lwrf42/emerge-dashboard

Note on naming

v0.1.0 publishes under @lwrf42/emerge-* (the maintainer's personal npm scope) because the @emerge org is not yet registered. The package contents and import shape (modulo the scope prefix) match what @emerge/* would be once that org is claimed.

emerge is a TypeScript harness for building durable, model-aware AI agents. Where most harnesses assume one model, one agent, one shot, emerge treats the model itself as a runtime variable — measuring its competence on the fly and adapting task structure to fit.

What makes it different

• Model surveillance. The harness continuously probes the active model's competence on the current task. When the gap is too wide, it automatically re-plans with finer decomposition or escalates to a stronger model. Weak models complete bigger tasks because the harness shrinks each step until it fits — across nested, opaque layers.
• Multi-agent as a primitive. Sub-agents, supervisors, and a streaming, bidirectional, addressable message bus are first-class kernel concepts. Topology is a value — supervisor / worker / pool / swarm / mesh / tree / pipeline / debate, nestable, with mandatory loop / recursion / mutual-respawn safeguards.
• Contract Custodian + Compliance Adjudicator. A dedicated agent role holds the master work contract verbatim through a pinned-context discipline that survives any compression. A separate evaluator gates completion against acceptance criteria. Children may negotiate quota mid-flight.
• AgentBlueprint composition. Specialized agents are assembled from typed slots, not subclassed — plug-and-play domain capability.
• Pluggable kernel. Provider, memory, tools, sandbox, telemetry, surveillance, modes, replay, experience, workspaces are swappable modules behind small, stable contracts. Wire your own; don't fork the harness.
• Long-horizon by default. Tasks are durable. Sessions resume across processes. Background work is a primitive, not a hack.
• Associative recall. Context is retrieved by semantic + structural + temporal + causal proximity — not last-N or naive grep — and every recall returns an explainable trace.
• Token-frugal by design. On-demand skill loading, layered compression, diff-only context updates, tool-result handles, and per-agent tool-result projections that strip / redact / cap / project before tokens hit the model.
• Operating modes + permission management. auto / plan / bypass / accept-edit / research / read built-in; user-defined modes pluggable; per-mode PermissionPolicy enforced at the kernel/sandbox boundary.
• Cost as a first-class observable. Per-call USD reported by providers; rolled up per agent / topology / contract; pre-flight forecast; cost ceilings enforced like any other budget.
• Honest reproducibility. Three tiers — record-replay (replays from log; fully reproducible), pinned (best-effort with logged divergence), free. No dishonest "same seed → same output" claims.
• Replay-grounded experience library. Every session is recorded; postmortem analysis distills Experiences keyed by problem-solving approach (not topic); surveillance reads them as priors at session start. The in-memory backend ships now; bundles are exportable / importable / mergeable for community sharing. The loop runs end-to-end today.
• Human-in-the-loop is a primitive, not a tool — human.request / human.reply / human.timeout envelopes; async approval queue.
• Workspace isolation. Each agent / topology branch / speculative branch gets an addressable Workspace (default: git worktree); merges are explicit, so parallel agents don't trample each other.
• Standard Schema everywhere. All contract boundaries use standardschema.dev-compatible refs — bring Zod, Valibot, ArkType.
• OpenTelemetry + W3C Trace Context end-to-end across nested agents.

Influences (and what we changed)

emerge studies — but does not copy — the public design conversation around Claude Code and similar harnesses. We borrow what's elegant (the loop, on-demand skill loading, sub-agent context isolation, polished TUI ergonomics) and reject what's constraining (single-vendor lock-in, single-agent assumptions, opaque kernel).

See VISION.md for the why and ARCHITECTURE.md for the how.

For AI agents

If you are an AI agent landing in this repo, start with AGENTS.md. For agent type-specific harness configurations, see docs/agents/. For SDK integration (embedding emerge in your own app), see docs/usage.md.

Try it with a real model

Run the hello-agent task (read README.md, write NOTES.md) against any of the three supported protocols. Each demo exits 0 with a "skipped" message when the env var is absent — safe in CI.

pnpm install && pnpm build

Anthropic (Claude):

ANTHROPIC_API_KEY=sk-ant-... node examples/hello-agent-anthropic/dist/index.js
# Optional: ANTHROPIC_BASE_URL=... ANTHROPIC_MODEL=claude-opus-4-7

OpenAI (GPT):

OPENAI_API_KEY=sk-... node examples/hello-agent-openai/dist/index.js
# Optional: OPENAI_BASE_URL=... OPENAI_MODEL=gpt-4o OPENAI_PROTOCOL=chat|responses

Any OpenAI-compatible service (Ollama, vLLM, llama.cpp, LM Studio, OpenRouter, your own):

EMERGE_LLM_BASE_URL=http://localhost:11434/v1 EMERGE_LLM_MODEL=llama3.2 \
  node examples/hello-agent-custom-url/dist/index.js
# Optional: EMERGE_LLM_API_KEY=... EMERGE_LLM_PROTOCOL=chat|responses

Status

M3c1 note: Inbox unification (AgentRunner consumes request envelopes addressed to it), supervisor LLM aggregation in supervisorWorker, Kernel.mountPostmortem() + auto-invoke in endSession(), OpenAI/OpenAI-compat providers, and three real-model demos shipped in this milestone.

M3a fix-up note: Adjudicator-gated session completion and pinned-context recall went live in the M3a fix-up commit. Mid-flight quota grant (bus-routed, applied atomically between preStep calls) and terminal-result envelopes on all failure/abort paths were also wired in the same pass.

The full roadmap with per-milestone deliverables and acceptance criteria lives at docs/design/roadmap.md.

Getting started

pnpm install
pnpm build
pnpm typecheck
pnpm test

Run all demos (skip-mode when env vars are unset):

for demo in hello-agent hello-mcp cycle-guard-trip eval-probes replay-smoke \
            topology-supervisor-worker weak-model-decomposition \
            hello-agent-anthropic hello-agent-openai hello-agent-custom-url; do
  node examples/$demo/dist/index.js
done

Contributing

The project is in its founding phase. The fastest way to contribute is to read the contracts in packages/kernel/src/contracts/ and open issues that challenge their shape — before implementations harden them.

License

MIT