Office Agent Harness Design

Definition

Office agent harness design is the product and infrastructure layer that lets an AI office agent turn workplace context, tools, files, connectors, model calls, and review loops into reliable task execution.

Current Synthesis

The source makes office-agent competition less about whether a product can list many skills and more about whether the harness can choose, compress, route, observe, and verify them. In Doubao Work, many skills, multiple entry points, file/context handling, and multi-agent execution show meaningful capability coverage, but 钟静伟 argues that missing harness details can raise cost and lower reliability.

The strongest design lesson is that office work inherits coding-agent harness problems while losing coding’s clean tests. Tool search, tool-schema compression, long-result compression, preloaded time and machine context, visible skill selection, and direct multi-agent delegation all shape whether the agent finishes a task efficiently or burns tokens through indirect coordination.

Key Claims

• Visible feature coverage is not enough; the harness decides whether tools and context become usable work.
• Skill abundance can become confusing when users cannot see what was loaded, truncated, or ignored.
• Multi-agent orchestration can add cost, latency, and information loss when delegation passes through unnecessary planning layers.
• Tool search, schema compression, result compression, and environment context are office-agent product quality issues.
• Model choice and BYOK are part of workflow fit, cost, and trust design rather than only advanced-user settings.
• Office agents need stronger review and observability because document, meeting, and workflow outputs are harder to test than code.

Evidence

• Skill and context handling: 272. 从飞书基座到Agent优先，豆包工作All in one紧追WorkBuddy reports that Doubao Work preloads many skills, shows context-share effects, and appears to truncate as more skills are added.
• Multi-agent architecture: 272. 从飞书基座到Agent优先，豆包工作All in one紧追WorkBuddy says Zhong Jingwei observed an architect-agent layer that dispatches subagents and then aggregates, which he argues may waste tokens and lose information.
• Missing optimizations: 272. 从飞书基座到Agent优先，豆包工作All in one紧追WorkBuddy names tool search, tool-schema compression, long-result compression, and basic time/computer context as expected harness optimizations.
• Competitive significance: 272. 从飞书基座到Agent优先，豆包工作All in one紧追WorkBuddy compares WorkBuddy as temporarily ahead at the harness layer while Doubao Work may catch up through ByteDance’s resources and Feishu context.

Counterevidence & Qualifications

• The observations come from a small number of reported product tests, so they may not represent every task or later versions.
• More explicit orchestration can still help if it improves safety, planning, permissions, or auditability; the source critiques unnecessary indirection, not all multi-agent structure.
• Office-agent harness quality cannot be separated from model capability, enterprise context, and customer permissions.

What Changed

• Added an office-specific harness concept grounded in Doubao Work and WorkBuddy product tests.
• Separated visible feature completeness from context/tool orchestration quality.

Related Concepts

• Agent Harness - broader model-external system that office-agent harnesses instantiate.
• Harness Engineering - engineering practice for building and improving harnesses.
• AI Office Agent - product category where the office harness operates.
• Multi-Agent Collaboration - orchestration pattern affected by delegation cost and information loss.
• Context Engineering - context-loading and compression layer inside the harness.
• Model Workflow Fit - model and surface fit that changes harness behavior.
• AI Inference Cost Structure - token and latency cost affected by harness design.
• Agent Permission Boundaries - safety and authority layer required for office execution.