Agents and what they can do

In Terere Harness, an agent is set up by a preset: which tools it can use, the instructions it follows and the skills it loads. This guide covers every capability an agent has, which ones you switch on, and how to create presets of your own.

Last updated 30 September 2026

Agents and presets

Every task runs with one agent preset, which the app also calls a mode. Four are built in, and any presets you create appear beside them under Custom.

The built-in agent presets and what each is for
Preset	Use it for
Standard mode	Everyday coding, file work and research. The agent calls tools directly and can use Skills, planning, sub-agents and workflows. Start here.
PTC mode	Batch work. The agent writes a small program that calls its tools in loops, which suits checking or summarising many files at once.
Minimal mode	A baseline with one shell tool and a fixed prompt, for comparing models or runs.
Creator mode	Adds capabilities to the app: plugins, and new agent presets of your own.

The capabilities in this guide depend on the preset a task uses:

Which capabilities each built-in preset includes
Capability	Standard	PTC	Creator	Minimal
Project instructions	Yes	Yes	Yes	No
Skills	Yes	Yes	Yes	No
Plan mode	Yes	Yes	Yes	No
Goals	Yes	Yes	Yes	No
Sub-agents	Yes	Yes	Yes	No
Claude Code and Codex agents	Once switched on	Once switched on	Once switched on	No
Workflows	Yes	No	Yes	No

Before you start

Open Settings, then General, and turn on Coding Tools. Without it the app hides the preset picker, and Creator mode can’t be started. You also need a model that supports tool calling; see Connect a model.

Create your own agent

Open Settings, then Agent presets.
Under Custom, choose Let the agent help me create a preset. This starts a new task in Creator mode.
Describe the agent you want: what it is for, what it may and may not do, and how it should report back. Give it a short name.
The agent writes the preset and asks your permission before installing it. Approve the install. Unless you have given it full access, it asks again each time it changes your setup.
Ask it to check the new preset loads. It then appears under Custom in Agent presets.

Some things you might ask for:

“Create a preset called "Reviewer" for code review. It should read files and run tests but never edit files, and it should always finish with a list of problems ordered by severity.”
“Create a preset called "Docs writer" based on Standard mode that always writes in British English, keeps paragraphs short, and checks every link it adds.”
“Make a preset like Standard mode but without the terminal, for people who only want to ask questions about the code.”

The more precisely you say what the agent must never do, the better the result. A preset controls which tools the agent has, so “never edit files” can be enforced by leaving the editing tools out, not just asked for in its instructions.

Use your agent

For one task: pick the preset in the preset picker when you start a new task. A task keeps the preset it started with.
For every new task: in Settings, then Agent presets, choose your preset’s card and Set as new task default.
To see what it contains: choose View configuration on its card. This shows the preset’s setup as read-only text.

Presets are saved on your computer, in your Terere Harness profile under %USERPROFILE%\.dsh, and apply to every task in that profile.

Change or remove an agent

To change a preset, start Creator mode again and describe what should change, naming the preset. To remove one, open the Plugins page from the sidebar, find the preset’s bundle, and remove it; or ask Creator mode to remove it for you. Removing a preset doesn’t affect tasks you already ran with it.

Project instructions (AGENTS.md and CLAUDE.md)

The agent reads instruction files from your project at the start of a task, so you can tell it how your project works once instead of in every message. It loads, in this order:

%USERPROFILE%\.dsh\AGENTS.md, for instructions that apply to every project.
AGENTS.md and CLAUDE.md in each folder from the project root down to the task’s working folder, broadest first. The project root is the nearest folder above the working folder that contains .git.
AGENTS.local.md and CLAUDE.local.md in the same folders, after the main files, for your own additions.

A CLAUDE.md with the same text as the AGENTS.md beside it is read once. When the agent reads or edits files in a deeper folder, that folder’s instruction files are added. Changes you make to the files yourself are picked up the next time the agent reads, writes or edits a file.

All instruction files together are limited to 64 KB. Over that, the broadest files are left out first, and the conversation shows a notice naming them. The @path imports and .claude/rules/ folders that Claude Code understands aren’t read.

Skills

A skill is a set of instructions for one kind of job, such as writing release notes or preparing a deployment. The agent sees each skill’s name and description, and loads the full instructions only when a task needs them. You can also start one yourself.

Where skills live

Put each skill in its own folder containing a SKILL.md file, under any of these folders:

.dsh\skills or .agents\skills in your project’s root folder, for one project.
%USERPROFILE%\.dsh\skills or %USERPROFILE%\.agents\skills, for every project.

A single file named after the skill, such as release-notes.md, also works directly in those folders. Skills in deeper subfolders aren’t found. New, renamed and deleted skills are picked up without restarting.

Write a skill

SKILL.md starts with a short header giving the skill’s name and description, followed by the instructions:

---
name: release-notes
description: Write release notes from the commits since the last tag.
---

Read the commits since the last tag with git log.
Group them into Added, Changed and Fixed.
Write one plain sentence per change.

The name uses lower-case letters, numbers and hyphens, such as release-notes. Write the description so the agent can tell when the skill applies. Two optional header lines change who can start a skill: disable-model-invocation: true means only you can start it, and user-invocable: false keeps it out of your / menu so only the agent uses it. A skill with a broken header, or a name in another form, is skipped and the conversation doesn’t say so, so check the header first if one doesn’t appear.

Start a skill yourself

Type / in the message box and pick the skill from the list, or type /release-notes directly, followed by anything else you want to say. The skill’s instructions are loaded for that message. Skills the agent loads appear in the conversation as a Skill row, which expands to show the Instructions. Skills only you can start are marked user-only in the list. If a skill has the same name as one of the app’s own commands, such as /plan, the command wins.

Plan mode

In plan mode the agent explores your project and writes a plan, and waits for your approval before it changes anything.

Type /plan, or /plan followed by the task, or choose Plan from the message box’s + menu. A blue Plan button appears next to the access mode.
When the plan is ready it opens in the right-hand panel under Plan review. Choose Approve to let the agent start, or Request changes and then type what should change; the agent waits for your message.
To leave plan mode without a plan, choose the blue Plan button or type /plan off.

Plan mode is guidance, not a lock: the agent keeps all its tools. If nothing may change while it plans, also choose a stricter access mode such as Read Only. Plan mode stays on after a restart until you leave it. Earlier plans stay available as cards at the end of the turn that produced them.

Goals

A goal is one objective the agent keeps working towards over many turns, for example “all tests pass on Windows”. While a goal is active and the agent is idle, the app gives it another turn, up to 256 turns, until the agent marks the goal complete or reports it is blocked. Each task has one goal at a time.

/goal followed by the objective sets a goal and starts work on it.
/goal on its own shows the current goal, how many turns it has used, and what you can do next.
/goal edit followed by new wording changes the objective.
/goal pause and /goal resume stop and restart the automatic turns.
/goal clear removes the goal.

You can also ask the agent in plain words to set a goal, and pause, resume, edit or clear it from the goal bar by the message box. The /goal commands run straight away and don’t use the model. Pause the goal before stepping away if you don’t want it to keep going. After you restart the app, an active goal waits until you type /goal resume.

Sub-agents

The agent can hand self-contained parts of a task to sub-agents, which run separately and report back. You don’t need to set anything up: ask for it (“split this into independent parts and use sub-agents”) or let the agent decide.

To limit how deep sub-agents can nest, how many can run at once, and which models they may use, open the Plugins page and choose Subagent.

Claude Code and Codex agents

If you use Claude Code or Codex, the agent can hand a task to either of them. The task runs as a real Claude Code or Codex session on your computer, signed in with your own account, and the answer comes back to your Terere Harness task.

Switch them on

Sign in on this computer first: for Claude Code, run claude in a terminal and follow its login; for Codex, run codex login. Terere Harness uses that sign-in, and never asks for or stores your Claude Code or Codex credentials.
Open the Plugins page from the sidebar and, under Official, turn on Claude Code agents, Codex agents or both. Both start switched off.
Start a new task. The agent now has a subagent_claude_code or subagent_codex tool in Standard, PTC and Creator modes. Tasks that were already open may not see it.

Then ask for it, for example “ask Codex to review the changes in this branch and list anything risky”. Turning the switch off removes the tool again.

What they can and can’t do

One task at a time, no follow-ups. Each hand-over starts a fresh session that gets one written task and runs until it finishes. It can’t be continued or resumed, so the agent has to put everything the task needs into that one message.
Unattended. Claude Code and Codex can’t ask you questions or for permission while they work. Claude Code refuses anything your Claude Code settings don’t already allow, so out of the box it can read your project and answer, but it can’t edit files unless your Claude Code permissions already allow edits. Codex doesn’t ask for approval and runs within its own sandbox settings, so anything the sandbox blocks fails.
Same folder, real changes. They work in the task’s working folder and can change your files directly. If the agent waits for the answer, stopping the task cancels the run; if it ran them in the background, ask the agent to stop the job. Changes made before cancelling stay.
Only the answer comes back. Your task gets their final answer, or an error if they fail. Their reasoning, commands and file changes don’t appear in your conversation, so ask the agent to check the files afterwards.
Their own settings and model. They use your Claude Code or Codex settings, including the model chosen there, and the work runs on that account. The Subagent limits on the Plugins page don’t choose their model.
Their own permission rules. With Auto review, the review checks the hand-over itself; once started, Claude Code and Codex follow their own permission rules, not your task’s access mode.

Workflows

For large jobs split across many sub-agents, such as auditing every file in a big project or researching a question from several angles, the agent can write a short script that starts sub-agents in parallel or in stages and combines their results. It only does this when you ask for a workflow or for large-scale orchestration; for one or two sub-agents it uses them directly.

The conversation shows each workflow run with its phases and members, and you can open a running member’s conversation. A workflow can also run in the background while the task continues; the job list at the top of the conversation shows its progress and can stop it. A cancelled or failed workflow reports an error, never a partial result. Workflows are available in Standard and Creator modes.

Features on the Plugins page

Some capabilities are separate bundles that you switch on or off on the Plugins page, under Official. A new installation of Terere Harness starts with these settings; if you installed an earlier version, your choices are kept and you can switch each one on yourself. If the page says the change takes effect at the next start, restart the app.

Optional bundles on the Plugins page and how a new installation starts
On the Plugins page	What it adds	New installation
Automation tasks	Run tasks in your conversations at a set time or on a repeating schedule.	On
Voice input	Dictate messages; speech is turned into text on your computer.	On
Auto Authorization Review	Adds the Auto review permission mode.	On
Agent Teams	A lead agent with named teammates and a shared task board.	Off
Claude Code agents	Let the agent hand tasks to Claude Code on this computer.	Off
Codex agents	Let the agent hand tasks to Codex on this computer.	Off

Agent Teams (experimental)

Agent Teams lets one lead agent create named teammates, send them messages and track work on a shared task board. It ships switched off. To try it, open the Plugins page and turn on Agent Teams, then ask for a team in a task, for example “use a team of two to review and then fix the failing tests”. The conversation shows the team’s members and task board. While Agent Teams is on, it replaces the regular sub-agent tools; workflows keep working.

Automation tasks

Automation tasks let the agent run an instruction later, once or on a schedule, in the same conversation. Ask for it in plain words, for example “every weekday at 9:00, check the nightly build and summarise any failures” or “in 20 minutes, check whether the deployment finished”. The agent can create, list, change and delete these tasks.

A task can run once after a delay or at a set date and time, at a fixed interval of at least a minute, daily or weekly at a local time, or on a cron schedule.
Automation tasks in the sidebar lists every task. Open one to change its name, instruction or run time, see its past runs, delete it, or go to its conversation.
A clock in the conversation’s header shows it has active tasks.
The agent is also told the current time and your time zone, so it knows when “tomorrow morning” is.

Terere Harness has to be running for a task to run. If it was closed, a repeating task runs once for the latest time it missed, not once for each. A conversation with active tasks can’t be archived until you stop them, which deletes them.

Auto review (experimental)

Auto review is an access mode beside Read Only, Workspace Write and Full access. Before each tool call, the same model your agent uses checks the action. Ordinary work inside your project goes ahead without asking. Actions such as deleting files that existed before the task, changing production systems, writing to outside services or changing security settings need your clear instruction, and sending sensitive data outside is always refused. When the review refuses a call, you’re asked to approve or reject it. In PTC mode each tool call inside the agent’s program is reviewed, but starting the program itself and anything the program does directly, without calling a tool, are not.

To use it, open the access mode picker in the message box (or type /permission), choose Auto review, marked EXP, and confirm the warning. /permission auto switches directly. It applies to the current task only; it can’t be chosen as the default for new tasks. It comes from the Auto Authorization Review bundle on the Plugins page.

Auto review is experimental. It runs without a sandbox, the model can allow something unsafe or refuse useful work, and every review uses extra tokens. Claude Code and Codex agents keep their own permission rules.

Voice input

Voice input turns speech into text in the message box. Speech is recognised on your computer by the SenseVoice model, which runs on the processor.

The first time, the app offers Go to setup. Choose it, then Download and prepare to download the speech model (about 240 MB). Nothing is downloaded before this. Model download source lets you choose where it downloads from.
Once it’s ready, choose the microphone between the model selector and Send, allow microphone access if Windows asks, speak, and choose Stop. The text appears in your message for you to edit. Cancel or Escape discards the recording.

In the Voice input bundle’s details on the Plugins page, set the language to English, Chinese, Cantonese, Japanese or Korean, or leave it on automatic. It needs a 64-bit Intel or AMD version of Windows; Windows on ARM hasn’t been tested. Switching Voice input off keeps the downloaded model on disk.

More tools with MCP servers

MCP servers add tools from other software, such as your issue tracker or a database. Servers you add apply to every task. See Add MCP servers.

If something doesn’t work

“Turn on Coding Tools in General settings”

Coding Tools is off. Turn it on in Settings, then General.

A preset shows “Failed to load”

Choose View configuration on its card to see the error, then start Creator mode and ask it to fix that preset, pasting the error. The other presets keep working in the meantime.

My new preset doesn’t appear

Check that you approved the install in the Creator task. If the task ended before installing, ask it to install and verify the preset.

The agent ignores part of its instructions

Instructions are guidance; tools are limits. If something must never happen, ask Creator mode to remove the tool that does it from the preset. Models with weak tool calling also follow presets less reliably; see Connect a model.

My skill isn’t in the / list

Check that the folder sits directly under one of the skill folders, that SKILL.md starts with a header containing name and description, and that the name uses only lower-case letters, numbers and hyphens. Also check the task uses a preset that includes Skills.

The agent doesn’t use Claude Code or Codex

Check the switch is on in the Plugins page, then start a new task in Standard, PTC or Creator mode. A preset you created yourself may have left the tool out; ask Creator mode to check.

A hand-over to Claude Code or Codex fails straight away

Make sure you are signed in on this computer: run claude or codex login in a terminal. Terere Harness doesn’t sign in for you, and the error it passes back doesn’t say why the sign-in failed.

An automation task didn’t run

Terere Harness must be open at the time. Open Automation tasks in the sidebar to check the task is still active and when it last ran.

The microphone does nothing

The speech model isn’t ready yet. Choose the microphone to see what’s needed, then follow Go to setup. If Windows blocked microphone access, allow it for Terere Harness in Windows’ privacy settings.