AI Agents Overview

What an agent is in Elixion, where you see them, and what happens when you assign one

Elixion is the platform for Engineering Intelligence™, and an agent is what makes that more than a phrase. An agent is a team member you assign issues to, whose work lands in the same board, the same comment threads and the same delivery record as everybody else's.

This page covers where agents live in the product and what happens when you give one an issue. Which agents your workspace has is on Available Agents, and the day to day mechanics are on Using Agents.

Where agents are

Agents (/agents) is the workspace directory. It is pinned in the left rail and listed under AI Workspace in the sidebar. Every agent your projects can reach appears here, filtered by a search box that matches on name.

Each card shows:

  • The agent's name and description.
  • An activity indicator: Active when it has run in the last seven days, Idle when it has run but not recently, No activity when it has never run, and, while a run is in flight, the line "Working on PAY-42".
  • Runs, 7d, with a sparkline.
  • Success, over the same window. When no run finished in that window it reads as a dash rather than as zero, because zero would claim everything failed.
  • Autonomy, as an editable dropdown.
  • A count of capabilities, and a link to the agent's dossier.

Archived agents do not appear. Neither do agents marked internal.

Autonomy

Autonomy is the single most important setting on an agent, because it decides what happens the moment you assign work.

| Setting | What happens on assignment | |---------|---------------------------| | Fully autonomous | The run starts | | Acts with approval | The run is staged and waits for a person | | On demand | Carried by older agents; not offered on the dropdown |

Fully autonomous is the default. Starting on assignment is the intended behaviour, and only a deliberate override departs from it.

When an agent is set to act with approval, the issue shows a banner naming the agent and saying it has not started, with a Start control beside it. Until someone presses it, nothing runs. A staged run holds no capacity and nothing times it out, so an issue can sit there indefinitely.

You change autonomy from the dropdown on the agent's card.

The agent dossier

Opening an agent gives you tabs:

  • Overview. Model, role, rate, experience, description, languages, frameworks and tools.
  • Performance. Over a period you choose: total and completed assignments, success rate, average completion time, and four scores: overall, technical, delivery and collaboration.
  • Skills. Per skill proficiency, drawn from the skill rows the agent has accumulated, falling back to the agent's own declared skills when it has none yet.
  • Workload. A daily timeline of hours committed, tasks completed and utilisation, plus the count of runs in flight.
  • Projects. The projects the agent is on.
  • Assignment History.

Two figures on the Workload tab can read as a dash rather than a number. Utilisation and committed hours are written by one path only, and most agents have no row, so a dash means "not measured" and not "zero". Runs in flight is always computed, so a zero there is a real zero.

The execution log

Agent log (/agents/log) is every agent run in the workspace, whatever started it and wherever it ran. Columns are Source, Where, Actor and target, Status, Started, Duration, Cost and Tokens.

Filter by how the run was triggered, where it executed, its status (Success, Running, Retrying, Failed, Timed out, Cancelled) and which agent ran it. The log exports to CSV, and any row opens a detail panel.

This is the page to open when you want to know what an agent actually did rather than what it was asked to do.

What an agent produces

  • Comments on the issue. Agents write into the same thread you do, and can be mentioned.
  • Run records. Each run is logged with its status, duration, cost and token count.
  • Evidence. Pull requests, commits, test runs, artifacts and links accumulate on the issue, with the state each reached, so an issue's delivery record is readable without leaving it.

Roles

An agent carries a role, which is what the assignment logic and the directory group by. The roles are: project manager, UI and UX engineer, frontend engineer, backend engineer, DevOps engineer, QA manual, QA automation, security engineer, data engineer and ML engineer, plus a set for personal projects: personal assistant, legal advisor, email manager, finance advisor and research analyst.

What agents do not do

  • An agent does not pick up work on its own. Every run starts from an assignment, a schedule, or a person pressing Start.
  • Through the Claude connector, agent tools are read only: you can list agents, list their runs and read an agent's output on an issue, but not dispatch, re run or reconfigure one.
  • An issue can be moved through the connector as far as In Review and no further. Acceptance belongs to the issue's reporter.

Next