--- title: MCP and skills made the agent harness more important than the model choice. publication: Agent Techniques Weekly slug: 2026-W21 issueNumber: 5 isoYear: 2026 isoWeek: 21 cadence: weekly publishedAt: '2026-05-23' periodLabel: Week 21 of 2026 canonicalUrl: https://brianletort.ai/industry/agents/2026-W21 schemaVersion: 2026.05.02 technique: name: Harness engineering mode: build agentCapabilities: - vendor: OpenAI product: Codex mode: build date: '2026-05-23' - vendor: Anthropic product: Claude / Claude Code mode: cowork date: '2026-05-23' - vendor: Microsoft product: Copilot / Scout mode: automate date: '2026-05-23' skillsAndConnectors: - ecosystem: MCP-capable agents name: Connector-backed workflow type: connector date: '2026-05-23' - ecosystem: Claude / Cursor / Codex / Hermes name: Reusable skills type: skill date: '2026-05-23' proofOfValue: - actor: Practitioner teams workflow: Coding, research, operations, and professional drafting evidence: confirmed scorecardAsOf: '2026-05-23' --- # MCP and skills made the agent harness more important than the model choice. *Agent Techniques Weekly · Issue 05 · Week 21 of 2026 · Published 2026-05-23* ## Big Read W21's read is that agent performance is increasingly shaped by harness quality: what tools the agent can call, which skills it can load, what data it can reach, and what policy gates constrain action. Model quality still matters, but a stronger model inside a weak harness underperforms a capable model with the right connectors, memory, and verifier. ## Technique of the Week ### Harness engineering Mode: `build`. Design the environment around the agent: tools, files, skills, permissions, memory, and checks. **Why it matters.** Harness engineering is how agent behavior becomes repeatable. It also creates portability: the same workflow can often move across Claude, Codex, Cursor, OpenCode, or other MCP-capable runtimes. **Anatomy:** - **Goal.** State the outcome and the boundary of delegated work. - **Context.** Give the agent the sources, files, examples, and constraints it needs. - **Tools.** Limit actions to the connectors, commands, and systems required for the job. - **Verifier.** Define how the output is checked before it is trusted. - **Escalation.** Name what requires human review, approval, or rollback. **Where it shows up:** - MCP connectors - tool allowlists - skill bundles - workspace permissions Source: [Addy Osmani, Agent Harness Engineering](https://addyosmani.com/blog/agent-harness-engineering/). ## New Agent Capabilities ### OpenAI — Codex Mode: `build`. Date: 2026-05-23. **Capability.** Workspace-bound coding assistance increasingly organized around issues, tests, review, and automation hooks. **Read.** The durable read is not any single coding model; it is the emergence of a build harness where context, tools, and verification are part of the workflow. Source: [Cursor MCP and background agent guide](https://laxaar.com/blog/building-ai-agents-with-cursor-1748650000028). ### Anthropic — Claude / Claude Code Mode: `cowork`. Date: 2026-05-23. **Capability.** Persistent instructions, skills, and agentic coding workflows point toward reusable professional work patterns. **Read.** Claude-style workflows are strongest when the human supplies policy, examples, and review criteria that survive beyond one chat. Source: [Claude Skills and MCP Servers practitioner guide](https://codersera.com/blog/claude-skills-mcp-servers-practitioner-guide-2026/). ### Microsoft — Copilot / Scout Mode: `automate`. Date: 2026-05-23. **Capability.** The Microsoft agent surface is moving toward M365-grounded action, background coordination, and enterprise controls. **Read.** The enterprise differentiator is governed access to mail, calendar, files, identity, and policy, not generic answer quality. Source: [Cursor MCP and background agent guide](https://laxaar.com/blog/building-ai-agents-with-cursor-1748650000028). ## New Skills and Connectors ### MCP-capable agents — Connector-backed workflow Type: `connector`. Date: 2026-05-23. **Signal.** Agents increasingly use connectors to reach source systems instead of relying on pasted context. **Why it matters.** Connectors turn chat into work by letting agents read the system of record and return traceable output. Source: [Claude Skills and MCP Servers practitioner guide](https://codersera.com/blog/claude-skills-mcp-servers-practitioner-guide-2026/). ### Claude / Cursor / Codex / Hermes — Reusable skills Type: `skill`. Date: 2026-05-23. **Signal.** Repeated workflows are being packaged as skills, rules, plugins, and templates. **Why it matters.** Skills are the portability layer for operating knowledge; they prevent teams from re-teaching the same workflow every week. Source: [Claude Skills and MCP Servers practitioner guide](https://codersera.com/blog/claude-skills-mcp-servers-practitioner-guide-2026/). ## Proof of Value ### Practitioner teams — Coding, research, operations, and professional drafting Evidence quality: `confirmed`. **Claim.** The strongest reported wins come from bounded workflows with clear checks, not open-ended autonomy. **Read.** Treat value claims as credible when the workflow, baseline, and verifier are visible. Treat broad percentage claims without methods as directional at best. Source: [Agent Harness Engineering](https://addyosmani.com/blog/agent-harness-engineering/). ## Enterprise Readiness - **permissioning.** Agents need least-privilege access scoped to the workflow, not broad user-equivalent authority by default. - **verification.** Every higher-autonomy workflow needs a deterministic check, source trail, rubric review, or human approval gate. - **auditability.** Background agents should produce inspectable logs of prompts, tool calls, approvals, outputs, and state changes. - **cost.** Loops need budgets and stop conditions because repeated agent calls can turn productivity experiments into runaway spend. ## Scorecard As of 2026-05-23. | Mode | Leading pattern | Representative tools | Control gap | |---|---|---|---| | chat | Structured context and critique loops | ChatGPT, Claude, Copilot Chat | Quality still depends on the user's review discipline. | | cowork | Human-supervised delegation with persistent project memory | Claude Cowork, Microsoft Copilot, Cursor | State, approvals, and source grounding must be visible. | | build | Workspace-bound agents with tools, tests, and worktrees | Claude Code, Codex, Cursor, OpenCode | Verification quality determines whether speed becomes rework. | | automate | Scheduled loops with state, tools, and escalation gates | Codex Automations, Microsoft Scout, Hermes, OpenClaw | Always-on agents need identity, audit trails, budgets, and stop conditions. | ## Try This ### Run a Harness engineering experiment 1. Pick one recurring weekly task with a clear definition of done. 2. Write a one-page loop contract: goal, context, tools, verifier, stop condition, and escalation rule. 3. Run it manually once with an agent and record where the verifier was weak. 4. Only automate the task after the verifier catches the most likely failure mode. **Expected outcome.** A reusable workflow contract and a clearer read on whether the task is ready for cowork, build, or automate mode. ## Watchlist - **Next 7 days — Copilot and Scout agent releases.** Microsoft's advantage is governed enterprise context; any new background or M365 action capability changes the automation surface. - **Next 7 days — Claude and Codex skill ecosystems.** Reusable skills and plugins are the leading indicator that agentic workflows are becoming products, not prompts. - **Next 30 days — Hermes, OpenClaw, OpenCode, and adjacent OSS harnesses.** Open-source harnesses reveal which control points matter most: memory, channels, terminal build loops, or automations. - **Next 30 days — Evidence-backed value claims.** The newsletter should elevate wins with named workflows, baselines, and verification methods, not generic productivity claims. ## Changelog - Backfilled Agent Techniques Weekly issue 05 for Week 21 of 2026. --- Source of truth: `src/data/industry/agents/2026-W21.ts`. Canonical HTML: .