Agent Farmbeta Join early access

District V of XNine things

Works as a team

Agents on different engines plan, split and check each other.

A team around a table while one of them sorts sticky notes on the wall

One agent can only do so much, and five of them in the same folder get in each other's way. Here a lead plans and a team carries it out, each on the engine that suits it, writers taking turns. Agents read each other's sessions and leave each other notes; you see everyone in one column, answer from one list, and your word overrides any of them.

11 Together

Missions · four engines, one team

A lead, a team, an honest verdict.

Give a mission to a lead. It plans round by round, hands tasks to agents on different engines, reads what comes back, and ends by saying how it really went. Partial is an answer too.

ROUND 1 / 8planning
managerlead · Claudeplanning round 1 developerClaudeidle reviewerCodexidle testerAntigravityidle writerOpenCodeidle
Verdictdonepartial

CSV and JSON lines export and pass their tests. The Windows path was never run, so the mission says partial, not done.

ConsoleAgent Farm — MissionsLive Timeline

Missions

Dynamic orchestration: a lead agent plans round by round, delegates to the team, and finishes with an honest verdict.

▾ SentryRUNNINGBacklog.md · item 4 of 11 · one clean session per item

A campaign takes the next item off a project backlog, opens a clean unattended session on it as an ordinary task row, and when that row ends it reads what the session left behind and opens the next one — until you stop it.

MissionLead + teamStatusRoundsTasksTokensCostStarted
›Timeline export, end to endADPT manager +3 running R3/87412.6k$0.62 (mixed)Oct 11, 09:12
›Q3 earnings, ten companies compared manager +2 done R4/66301.2k$0.48 (mixed)Oct 10, 16:40
›Do two agents really argue?ADPT reviewer +1 donepartial R2/5172.3k$0.10Oct 9, 17:53
›A chaired council that really arguesADPT manager +0 cancelled R1/8358.9k$0.11 (mixed)Oct 9, 21:17

03 The console

Rebuilt from real screenshots · demo data

This is the screen you work in.Take it apart.

The real Agent Farm Console, rebuilt. Scroll and it assembles: the leg of the chain, who is waiting for you, a session that rebuilt the extension and kept talking, the context meter and the quota shield.

Your VS Code theme
FileEditSelectionViewGoRunTerminalHelpUntitled (Workspace)—▢✕
456
AGENT FARM
⌄ AGENTS
analyst 2h ago
assistant 49d ago
⚠ developer STALLED ×4 · 220.9M tok
editor 3d ago
manager 15h ago
scholar 8d ago
writer 1d ago
youtuber 20m ago
⌄ TASKS
Plan 3/15 on · 8 manual
Market brief analyst · daily
Weekly episode youtuber · Fri
Weekly research scholar
› MISSIONS
⌄ TOOLBOX
Skills 2 user · 0 agent
MCP Servers 4 user
Hooks 9 user
⌄ LIVE FEED
Console (2)Agent Farm — Live TimelineAgent Farm — Now

Console

developer · working

DashboardProjectsMobileLivingNotices (1)QueueLive TimelineAgentsTasksMissionsStudioWikiToolbox
+ New session⇅
● 2 waiting for youlongest 6m
developer(#3efb20bf)Claude
acct-c⑂ 921.8M tk
VIDEO PIPELINE BUILD
working · 1 turns · ctx 110.2k / 1.0M
PRS.AgentFarm
youtuber(#ses_a61c)North-Mini-Code
OpenCodefree model
Run episode 12 through the video pipeline: script, voice-over, footage, cut, thumbnail…
working · 3 turns · ctx 48.2k / 256k
scholar(#7d2e51b0)Gemini
⛓ step 2 of 3640k tk
Chain: clean the survey data, run the chapter 3 regressions, then write Table 3…
your turn · 6m · 4 turns
analyst(#c90f3a72)Codex
Morning market brief: screen my watchlist for earnings, gaps and unusual volume…
your turn · 5m · 2 turns
writer(#1b84e6d3)Kimi K2
Claudeanother brain, same engine
Draft chapter 7 from the outline, in my voice, about 3,000 words…
working · 1 turn · ctx 82.4k / 1.0M
editor(#e52a7f18)Antigravity
Cut interview-take-4 into four shorts with burned-in captions…
done · 12m ago · 9 turns
› Resume an existing session
✎ Edit test/pipeline/editCompose.test.ts
Message developer… (Ctrl+Enter sends)
developerClaudeopus ▾xhigh ▾acct-c ▾bypassPermissions ▾Send after turn⏸ InterruptSend into turn
⑂ agent master*1 active · $41.20 today◷ 2h 40m · 91% / wk 17%⑂ 92ProjectGit GraphAgent Farm 3.931.0-beta
Reloading window · installing 3.932.0-beta
the session keeps its seat
⑂ 92+Legs of one relay chain
2 waitingNobody waits unseen
Same seatRebuilt itself, kept talking
Context meterFresh mind every leg
5H / weekQuota shield rotates accounts
Exploded view · the real Console, rebuilt
V·1

Missions

A lead that cannot write plans round by round and hands the work to a team on different engines, one writer per folder, until the goal is met or one of five caps is reached.

A mission, live

What it solves

Some work does not fit in one agent: it needs planning, doing, reviewing and testing, and the next step depends on what the last one found. Drawing that as a fixed chain in advance means guessing.

How it works · pick a step

An adaptive mission takes an objective with criteria instead: independent tasks run side by side in waves, a reviewer is handed the real output of the task it reviews, and agents can hold a conversation in which each turn reads what the others actually wrote.

V·1A goal and a team1 / 5
V·2

Shared sessions

Any agent can open any session by its id, whichever engine wrote it, see what was and wasn't done, and carry on.

What it solves

A session written by Claude Code is a file Codex cannot read, and the other way round. Work started by one engine stopped there, unless you copied the important parts across by hand.

How it works · pick a step

So the reviewer on Codex reads what the developer on Claude actually did, turn by turn, and carries on from there.

V·2Where it is1 / 5
V·3

Agent notes

Agents leave each other notes and hand tasks over, with a guard against loops.

What it solves

Agents find things the next agent needs, a locked database, an API that changed, a result, and they find work they cannot do themselves. Telling you every time turns you into the postman.

How it works · pick a step

Agents look at the board when a session starts: a2a.status gives unread notes, tasks they could take and leases they hold in one call. A relay's handover shows there too, as a card that resolves itself when the next leg starts.

Settings · 1
SettingDefaultWhat it does
agentFarm.a2a.preferredGraceHours2How long a task waits for the role it names before another agent may take it.
V·3A note1 / 5
V·4

Steer and veto

Your steer, veto or constraint outranks everything an agent was told, and reaches a running task with its next heartbeat.

A mission, live

What it solves

An agent on the wrong track keeps going until somebody stops it. Stopping it and starting over loses what it already got right.

How it works · pick a step

V·4Steer, veto, constraint1 / 5
V·5

Console

Every session in one column: who is working, whose turn it is, which leg of which chain.

The Console, live

What it solves

Every CLI has its own terminal and its own habits. With eight sessions running at once you cannot tell which is working, which is stuck and which is waiting for you.

How it works · pick a step

Your install, your subscription: the Console runs the CLI you already have, and the transcript lands where it always does, so every other page keeps reading it. Files and screenshots drop or paste straight into the box, and saved prompt snippets are a right-click away.

V·5Five engines, one card1 / 5
V·6

Countdown questions

An agent's question picks the suggested answer when time runs out. A permission request always waits for you, with every stuck session, in Needs You.

Waiting for you, in the Console

What it solves

An agent asks a question at midnight, nobody is at the desk, and the whole night goes by waiting for one click.

How it works · pick a step

The agent is told the timer answered, not you, and the transcript marks each answer as yours or the recommended one. The clock runs in VS Code itself, so it keeps going with the Console closed.

Settings · 1
SettingDefaultWhat it does
agentFarm.console.autoAnswer.timeoutSeconds300Seconds before a question picks its recommended answer. 5 to 7200.
V·6The countdown1 / 5
V·7

Notices and the queue

Agents report, remind and ask on one board, by level, and your answer becomes work in the queue. Your side of it: leave a job for an agent, now or when its group fills.

What it solves

Agents have things to tell you that are not urgent, and you have small jobs for them that come to mind while you work on something else. Both get lost in chats.

How it works · pick a step

Needs You only holds what stays stuck if you never touch it. News, a blown budget or a proposal, goes to Notices, which is made to be read later. An agent that asks there can sleep until you answer, spending no tokens while it waits.

Settings · 1
SettingDefaultWhat it does
agentFarm.noticeStaleHours24When an unread important notice or an unanswered question turns the bell into a warning.
V·7One row each1 / 5
V·8

Live Timeline

Every running session on one axis. Jump to one tool call among thousands; a stall raises a flag.

What it solves

Thousands of tool calls across a dozen sessions cannot be read as text. You need to see who did what, when, and where something stopped.

How it works · pick a step

Settings · 2
SettingDefaultWhat it does
agentFarm.stuckThresholdSeconds240Seconds of silence before a session is marked stalled. 0 = never.
agentFarm.notifyTurnEndoffA notification every time any agent finishes a turn. Can be noisy.
V·8A lane per session1 / 5
V·9

Worktrees

Parallel jobs work in separate git worktrees, so they never step on each other.

What it solves

Two agents in the same repository at once overwrite each other's files and mix their changes into one diff nobody can review.

How it works · pick a step

Settings · 1
SettingDefaultWhat it does
agentFarm.maxConcurrentPerAgent1How many of one agent's runs go at once; the other worktrees wait.
V·9Plan first1 / 5

Questions

Can two agents edit the same files at once?

Not by accident. In a mission with a working folder, writers take turns and only read-only steps share a wave. For jobs that should run side by side, each gets its own git worktree and branch.

Can a mission run away with my plan?

No. Every mission has five caps, cost, rounds, tasks, minutes and tasks at once, and when one is spent the lead may only finish. Repeating the same work to the same agent is refused.

What if nobody answers an agent's question?

When the countdown ends, the answer the agent recommended goes in, and the agent is told it was the timer. A destructive question and a tool permission are never answered on a clock: they wait in Needs You.

Can agents on different engines work on the same thing?

Yes. A mission's team can mix Claude Code, Codex, Gemini CLI, Antigravity and OpenCode, and any agent can read any engine's session in the same words.

Do agents talk to each other?

Through notes and tasks on the A2A board, and in an adaptive mission through a conversation in which each turn is shown what the others actually wrote. Nothing reaches you unless it is for you.

What is the difference between Notices, the queue and Needs You?

Notices is what agents tell you. The Work Queue is what you give them. Needs You is only what stays stuck until you act: a permission, a question no timer answers, a stalled session.