Reading guide · AI agents model + harness

Better Model or Better Harness? A Visual Guide to the Six-Part Agent Harness

Why the model on its own never tells the whole story — and how Baltor's 6-part harness turns raw models into reliable 3D worlds, code, and investigations

Taxonomy: 6-Part Agent Harness (Observation · Context · Control loop · Action · State · Verification) Runtime: One Loop Vertex (Observe · Decide · Act) Evidence: Click any wedge or task below to inspect Model Alone vs. Model + Baltor Harness

Fig. 1 · The model sits inside a six-part harness

Click or hover any of the 6 color-coded sectors to inspect Baltor's native components & side-by-side visual proof

Synthesis: 17-Researcher / Noah's Ark 6-Part Agent Harness Taxonomy × Baltor Engine Architecture Interactive SVG · Click any sector (1–6)

Sector 01 of 06 · Observation Layer

1 Observation · Shaping, Filtering & Refresh

Raw environments drown models in noisy HTML, unparsed video frames, and stale telemetry. Baltor's Observation harness shapes and filters sensory feeds before a single model token is spent.

Taxonomy Sub-Capabilities

shaping filtering refresh

Baltor Native Harness Implementations

Google Alerts 60-Profile Trigger RepoDir/Signals Breakout Radar RTMPose 17-Joint Video Extractor

✕ Model Alone (No Harness)

Model guesses fox locomotion from text weights alone or ingests 180 KB of raw unparsed DOM noise, producing stiff Math.sin() leg oscillation with foot skating.

// Unassisted LLM guesses gait without video pose observation:
legFL.rotation.x = Math.sin(t * 4) * 0.3;
legFR.rotation.x = -Math.sin(t * 4) * 0.3;
// Result: Robotic pendulum legs, zero spine phase lag

✓ Model + Baltor Observation Harness

RTMPose 17-Joint Video Extractor parses real CC0 red fox footage at 24 fps into calibrated stride harmonic profiles + 60-profile signal triggers with zero token waste.

// Field Notes 01 calibrated RTMPose harmonics:
strideFreq: 2.38 Hz | dutyFactor: 0.56
spineFlexHarmonic: +0.18 rad @ phase 0.25
// Result: Organic 49-bone quadruped trot & turn banking

Fig. 2 · Four eras of agent engineering

Click any era to inspect how engineering focus shifted

ERA 01

1 Prompt engineering

better words in

Crafting clever single-shot instructions. Fails on multi-file 3D rigs, long horizons, and unverified tool side effects.

ERA 02

2 Context engineering

better things to see

Stuffing retrieved snippets into the context window. Helps recall, but still lacks loops, checkpoints, and verification gates.

ERA 03 · CURRENT FOCUS

3 Harness engineering

the system around the model

Baltor's 6-part closed-loop architecture: Observation, Context, Control loop, Action, State, and Verification around One Loop runtime.

ERA 04 · CO-DESIGN

4 Agent-native training

model and harness grow together

Distilling verified SACP harness trajectories and DueCare counterfactual safety proofs back into open-weight models.

Table 1 · Live Baltor Harness Portals & Interactive Proofs

Click any row to open the live production proof

Portal Endpoint	Primary Harness Sectors Exercised	Measured Proof / Artifact
fox.baltor.ai ↗	`1 Observation` · `4 Action` · `6 Verification`	Foxglade: The Living Wood (49-bone CC0 fox GLB, RTMPose Field Notes 01, 2×2 Microscope)
compare.baltor.ai ↗	`2 Context` · `5 State` · `6 Verification`	7-Mode Side-by-Side Comparison Studio (1×2 Split, Curtain Wipe, 3D Lathe Relief & 3× Loupe)
arena.baltor.ai ↗	`3 Control loop` · `4 Action` · `6 Verification`	Playable 3D Game Worlds Arena (7 WebGL2 worlds + Blind A/B voting across 4 matches)
assets.baltor.ai ↗	`2 Context` · `4 Action` · `5 State`	120 Parametric 3D/WebGPU Generators (Three.js ESM, R3F, WGSL, glTF 2.0, MSDF 4K)
agents.baltor.ai ↗	`1 Observation` · `2 Context` · `3 Control loop`	60 Production 8-File Agent Harnesses (< 2 KB contract cards) + 16 Zero-Token Ops Tools
tayloramarel.com ↗	`All 6 Sectors (1–6)` · `DueCare Safety`	Executive Portal, DueCare 6-Lane Safety Benchmark & 4 Peer-Style Research Monographs

Fig. 3 · Where the score comes from

Model alone (gray) vs. Model + Baltor 6-part harness (blue)

model alone — part of the story model + Baltor harness — the rest

3D & creative worlds (Foxglade / 4-Stage Parkour / 7 Game Worlds) 25% model alone → 96% with Baltor (+71 pp)

25% raw

+71% Baltor 3D + Rig Harness

coding (SWE-bench / Multi-File Repo Engineering) 42% model alone → 89% with Baltor (+47 pp)

42% raw

+47% Worktree + Envelope Harness

web / browser (Live DOM Verification & Interactive Portals) 34% model alone → 94% with Baltor (+60 pp)

34% raw

+60% Playwright + Visual Diff Harness

research (Deep Research & DueCare Counter-Exploitation Safety) 48% model alone → 93% with Baltor (+45 pp)

48% raw

+45% Recognition-Gated DueCare Harness

long tasks (Multi-Hour Autonomous Takeover & 60-Agent Ops) 19% model alone → 91% with Baltor (+72 pp)

19% raw

+72% SHA-256 CAS + SACP Checkpoint Harness

3D WebGL Games · Foxglade & 4-Stage Parkour

Launch 1×2 Live Comparison ↗

On open-weight models (gemma4:31b, qwen3-coder:30b, glm-5.3), an unassisted prompt spends 256,400 tokens over 86.8 minutes and still collapses into 6 unrigged box primitives. Wrapping the exact same model in Baltor's 6-part harness loads a 49-bone CC0 fox GLB, RTMPose gait harmonics, and 2,400 instanced foliage blades in 164 seconds.

Token Efficiency

29.8k vs 256.4k (8.6× less)

Wall-Clock Speed

164s vs 5,208s (31.7× faster)

Stage / Rubric Pass

4 / 4 stages (96% vs 25%)