the harness

UnieAI Agent Core — a better harness for open models

Agent Core is the harness that makes open models smarter and the runtime that makes them cheap to run. It decouples session, sandbox, filesystem, orchestration and tools into efficient services — so hundreds of agents run in one process, not dozens of VMs.

Agent run

Analyze last quarter's complaints by product line

Query the revenue database
Read the complaints folder (1,204 files)
Call the logistics APIharness repair · auto retry
Go deeper: break down by product line
Deeper still: cross-check delivery delays
Done · report with sources attached

Same model — AIME 78.3% → 97.2%

+18.9pt

AIME 2025 uplift on MiniMax-M2 (78.3% → 97.2%) with Agent Core 2

100s–1,000

Concurrent agent turns per replica vs. ~10–40 for sandbox-per-agent*

~3 MB

Memory per agent turn — ~0 CPU while waiting on model & tools*

0s

Cold start — stateless, linear horizontal scaling

A better harness makes open models smarter

The same open model gets materially stronger and more reliable when it runs inside a purpose-built harness — better planning, better tool use, stable loops.

  • Agent Core 2 lifts MiniMax-M2 on AIME 2025 from 78.3% to 97.2%
  • Stable agent loops on open-source models
  • The same harness powers UnieAI Chat, Code & Studio
OpenmodelPlanningTools · MCPSandboxMemoryAgentic RAGAGENT CORE HARNESSAIME 202578.3%97.2%+18.9pt

Two halves of one agent-inference engine.

UnieAI Agent Core

CPU efficiency

Decoupled, async agent runtime — hundreds of concurrent turns per process, single-digit MB each.

UnieInfra

GPU efficiency

Token-efficient throughput density and low TTFT for agent inference. Converging with Agent Core into one engine.

How the harness is designed

Agent Core 1.0 — a production-grade native function-calling agent loop, hardened inside UnieAI Studio. Zero runtime dependencies.

Native tool loop

One iteration, one model step, driven by native tool_calls; results re-enter message history. Bounded concurrency, per-tool timeouts, and automatic repair of malformed calls.

Never gets stuck

Doom-loop guards (identical calls, same-tool streaks), transient-error retry with backoff, context-overflow recovery — and a soft-landing wrap-up that guarantees a text answer at the step budget.

Long-run context management

Zero-LLM compaction at entry plus mid-loop pruning of old tool outputs — long runs stay inside the window, with no compaction pause.

Tiered memory

The model edits its own Tier-1 core memory via a memory tool (char-capped buckets, threat-scanned writes); a background review agent reflects after replies.

Skills on demand

load_skill injects skill bodies only when needed, keeping the base context small.

Sub-agent forking

A task tool forks an isolated child loop and returns only a distilled conclusion to the parent.

Domain-agnostic — tools are injected

The core ships only generic tools (todo / memory / skill / task); your domain tools — search, SQL, APIs, GPU kernels — are injected as domainToolBuilders. Two production consumers prove the seam: UnieAI Studio injects kb, sql, bi, api and mcp; Kernel-Design-Agent injects CUDA-kernel tools. Zero runtime dependencies.

todomemoryskilltaskkbsqlbiapimcpCUDA

Hundreds of agents per process — not dozens of VMs

Traditional frameworks give every agent its own sandbox: hundreds of MB to GB each, seconds to cold-start, so high concurrency means spinning up hundreds of VMs. Agent Core decouples the agent into efficient services and multiplexes I/O-bound turns in a single process.

  • ~2–4 MB per turn; ~0 CPU while waiting on model & tools
  • No cold start — turns spin up instantly
  • Stateless replicas scale linearly: 10 replicas ≈ 2,560 concurrent turns
SANDBOX-PER-AGENT1 agentVM · ~1 GB1 agentVM · ~1 GB1 agentVM · ~1 GB~10–40 concurrent · sec cold-startvsAGENT CORE100s–1,000 concurrent · ~3 MB each

Build agents on a harness you can trust.