Agent Techniques Weekly
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
Design the environment around the agent: tools, files, skills, permissions, memory, and checks.
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.
State the outcome and the boundary of delegated work.
Give the agent the sources, files, examples, and constraints it needs.
Limit actions to the connectors, commands, and systems required for the job.
Define how the output is checked before it is trusted.
Name what requires human review, approval, or rollback.
OpenAI / build
Workspace-bound coding assistance increasingly organized around issues, tests, review, and automation hooks.
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.
Anthropic / cowork
Persistent instructions, skills, and agentic coding workflows point toward reusable professional work patterns.
Claude-style workflows are strongest when the human supplies policy, examples, and review criteria that survive beyond one chat.
Microsoft / automate
The Microsoft agent surface is moving toward M365-grounded action, background coordination, and enterprise controls.
The enterprise differentiator is governed access to mail, calendar, files, identity, and policy, not generic answer quality.
connector / MCP-capable agents
Agents increasingly use connectors to reach source systems instead of relying on pasted context.
Connectors turn chat into work by letting agents read the system of record and return traceable output.
skill / Claude / Cursor / Codex / Hermes
Repeated workflows are being packaged as skills, rules, plugins, and templates.
Skills are the portability layer for operating knowledge; they prevent teams from re-teaching the same workflow every week.
Evidence: confirmed
The strongest reported wins come from bounded workflows with clear checks, not open-ended autonomy.
Treat value claims as credible when the workflow, baseline, and verifier are visible. Treat broad percentage claims without methods as directional at best.
Agents need least-privilege access scoped to the workflow, not broad user-equivalent authority by default.
Every higher-autonomy workflow needs a deterministic check, source trail, rubric review, or human approval gate.
Background agents should produce inspectable logs of prompts, tool calls, approvals, outputs, and state changes.
Loops need budgets and stop conditions because repeated agent calls can turn productivity experiments into runaway spend.
Expected outcome: A reusable workflow contract and a clearer read on whether the task is ready for cowork, build, or automate mode.
Next 7 days
Microsoft's advantage is governed enterprise context; any new background or M365 action capability changes the automation surface.
Next 7 days
Reusable skills and plugins are the leading indicator that agentic workflows are becoming products, not prompts.
Next 30 days
Open-source harnesses reveal which control points matter most: memory, channels, terminal build loops, or automations.
Next 30 days
The newsletter should elevate wins with named workflows, baselines, and verification methods, not generic productivity claims.