AGENT TO OUTCOME / FIELD NOTES / OCTOBER 22, 2026

Choose the workflow.
Then choose the agent.

The most useful starting question is the job you need done. A clear outcome, usable context and a review step make the tool decision much easier.

A team preparing weekly client updates could start by comparing a dozen agent products. It could also start with a much more concrete brief: “Turn approved project notes into a one-page update, preserve the source of each claim, identify missing information and stop before sending.” The second approach creates something the team can test.

The names will keep changing. The work map is what lets you move between tools without rebuilding your understanding of the task each time.

Separate the model from the working environment.

A language model produces responses from the information it receives. An agent system adds a working loop: it can choose a step, use an available tool, inspect the result and continue within its instructions and controls. The surrounding software is often called an agent harness. It manages things such as tools, context, permissions, execution and logs.

A workflow specifies how work moves. Some steps can be fixed, such as validating a required field. Others can use an agent’s judgment, such as deciding which approved sources answer a research question. Use the simplest combination that produces a dependable result. Anthropic’s guide to effective agents explains the distinction between predefined workflows and more dynamic agent control.

Choose by the task and the place it happens.

Product categories overlap, so treat this as a starting map. Verify availability, account permissions and the particular tools your task needs. It is not a popularity ranking.

ChatGPT Work is a workspace for longer, multi-step work. Codex and Claude Code are useful choices when the task benefits from development tools, but product boundaries are expanding; do not assume a coding agent can only write code.

Claude’s terminology is also changing. Its official Cowork update describes chat and Cowork coming together, with the experience depending on rollout and plan. Claude Design has its own changing entry points. Check the dated Agent to Outcome tool guide when selecting a workshop tool.

Write the delegation brief.

Give the agent enough structure to succeed and the reviewer enough structure to judge the result. Our workflow planner captures seven essentials:

1. Goal: the decision or responsibility this work supports.
2. Deliverable: the output, format and definition of done.
3. Context: the audience, constraints, examples and relevant history.
4. Inputs: the approved sources and how to handle missing or conflicting information.
5. Access: the tools, data and actions the agent may use.
6. Human checkpoints: the person who reviews, approves and handles exceptions.
7. Evidence: the tests, costs and observations that show whether it helps.

For the weekly update, the first version might read approved notes and draft a document. Sending the message stays with the account owner. That boundary lets the team learn about quality before adding a consequential action.

Connections need a purpose and a boundary.

Plugins, APIs and MCP servers can provide capabilities or context, depending on the product. MCP is a protocol for connections between an AI application and servers that expose tools and resources; it does not make every connection safe or authorize every action. The MCP architecture guide describes the host, client and server roles.

Connect the smallest useful set. Distinguish permission to read a CRM record from permission to update it. Specify what happens when a source is unavailable, a tool requests broader access or retrieved content contains instructions unrelated to the task. Source material should be treated as evidence to assess, not authority to change the assignment.

Test the result and the route to it.

Run a few representative cases: an ordinary week, incomplete notes, conflicting dates and a tool failure. Inspect the final document and the relevant execution history. Check whether the agent invented a status, omitted a blocker, used an unauthorized source or continued past the review boundary.

Keep the input, expected behavior, observed result and revision together. Repeat those cases when you change instructions, tools or models. Anthropic’s agent evaluation guide is useful background on testing both outcomes and the paths agents take.

Measure total effort: setup, preparation, execution, review, correction and maintenance. For API workflows, include every model call, subagent, retry, tool charge and compute cost. Compare cost per accepted result, not just the token cost of the first draft. A cheaper model that needs repeated correction may be more expensive for the whole job.

Add orchestration when the job earns it.

A single agent is a sensible first pilot. Specialist subagents can help when work divides into independent tasks with clear inputs and outputs. For example, separate research assignments can return evidence to one coordinator. They also introduce coordination, review and usage costs.

Give each specialist a limited role, required output and budget. Decide who resolves disagreements, which evidence must survive the handoff and who owns the final result. More agents create more activity; your acceptance criteria determine whether they create more useful work.

← Back to Agency Field Notes