Card 02 · AI agents, explained
What is an agent harness? The model is the engine, the harness is the car
An agent harness is everything around the model that lets it act: the loop that picks the next step, the tools it may call, the memory it reads, the limits it runs under and the record of what it did. The model supplies the thinking; the harness supplies the hands, the memory and the brakes.
01 — The loop
A model predicts text. The harness makes it work.
Ask a model for Friday's report and you get a good guess with no hours, no clients and no dates. The harness is the loop around it: pick the next step, use a tool, read what came back, decide again, and keep a record of what it did. Priya is not typing any of it.
02 — Memory
It remembers Friday on Monday.
Memory is what the agent carries from one run to the next: what was decided, what changed, which client is which. Without it, every Monday begins with months of threads to re-read. Priya's agent opens Monday already knowing the invoice moved to net-30 last week.
03 — Skills
The method is written once, then followed.
A skill is the written method the agent follows: when to use it, what to read, the steps to take, how to check the result. Priya wrote hers in March, decisions first and two pages at most. Every Friday since has come back in that same shape.
04 — Tools and MCP
MCP is the plug. The tool is what's on the other end.
Tools are how the agent acts outside the conversation: read the timesheet, check the calendar, post the draft. MCP is one standard connector, so the same harness reaches accounting, chat and calendar without a build for each. A fourth tool becomes a connection, not a project.
05 — Permissions
It can draft anything. It sends only when you say so.
Permissions set what the agent may do alone, action by action: read, draft, send, delete. Priya's agent may read timesheets and write the draft; sending to eight clients asks her first. That rule is why she can leave it running on a Friday afternoon.
06 — Routines
Friday at two, nobody asked.
A routine runs the harness on a schedule or a trigger instead of a prompt. Priya's fires Friday at two, then Monday morning with the weekend's changes folded in. The report is ready before anyone remembers it is due.
07 — Channels
It lands where your team already talks.
Channels are where the agent shows up: a chat thread, an email draft, a file in the shared space. Nobody learns a new app to read the report. Priya's draft appears where her firm already works, and the client copy leaves as email.
08 — People
Every agent has a person's name on it.
The last part is not software. An agent acts for one named person and its work carries that name, so a wrong number has an owner. When Priya is out, Leo's agent picks up Friday's report because the work sits in the shared space, not in her chat history.
All seven parts in place. Priya's report is ready on Friday at two.
Tools for this
Every price and licence below is what the tool’s own page said on 18 November 2026, with two sources behind each record. There are no scores and no order of preference; where TalentOS is one of the tools it is marked ours and listed like the rest.
A terminal pair-programming agent that edits files in your git repository and commits each change with a written message.
For a developer who works from the terminal and wants every agent edit to arrive as a reviewable commit.
$0 (free) · Open source, Apache-2.0
An AI code editor whose agent edits across files, runs commands, reviews its own changes and can run tasks in the cloud.
For a developer who wants an agent inside the editor they already use, with team rules and skills shared across the codebase.
$8 per month (seat) · Closed source
GitHub's coding agent, in the editor, the CLI and on github.com, that can pick up an issue and open a pull request against the repository.
For a team that already runs its work on GitHub and wants the agent where the issues, pull requests and reviews live.
$10 per user per month (seat) · Closed source
An open-source agent that runs on your own machine as a desktop app or in the terminal and extends itself with MCP servers.
For a developer or analyst who wants a local agent pointed at their own tools and their own model keys.
$0 (free) · Open source, Apache-2.0
An open-source coding agent you can run yourself or use on the vendor's cloud, with a control centre for keeping several agents working on their own.
For a team that wants to self-host its coding agents and keep its own model key.
$0 self-hosted (free) · Open source, MIT
Anthropic's coding agent that runs in the terminal or an IDE and works on a real repository: it reads files, edits them, runs commands and commits.
For a developer or a small engineering team that wants an agent working inside its own codebase, not in a chat window.
$20 per month (seat) · Closed source