Several AI windows can still leave you doing all the coordination. AgentRelay 2.11.0 connects existing conversations so an agent can receive a task, execute it independently and report back to the controller. Your next step becomes reviewing results instead of carrying messages between windows.
A team needs a handoff, not just more chat windows
Imagine a bug fix: Codex owns the scope and integration, Antigravity changes an agreed set of files, and DeepSeek Harness reviews an independent piece of logic. Each worker receives the goal, its boundaries, the expected evidence and the controller’s current conversation address. Workers save their results and proactively send DONE or BLOCKED.
The controller checks the report, validates the work and chooses the next action. Independent tasks can move forward together; dependent tasks wait for the required result. Communication can run in parallel, while branches and file ownership still protect the repository. Direct messaging does not make simultaneous edits to the same file safe.
Where the savings and speed can come from
Reuse compatible clients and subscriptions you already have, subject to their terms and limits. Route a bounded implementation or review task to an appropriate engine and reserve more demanding reasoning for integration and difficult decisions. AgentRelay is the local communication layer; it does not include an AI subscription or remove account quotas.
Less copying, repeated briefing and status checking can reduce coordination work. A controller can do independent work while a worker runs, then react to the returned result. This is a practical cost-control strategy, not a measured savings percentage. Model usage, subscription coverage and the amount of duplicated context determine the actual bill; messages and reports can still consume tokens.
Move the next step from polling to completion events
With polling, the controller repeatedly asks whether a worker is done. With proactive reports, the worker that produced the result sends it back. The normal loop becomes assign → execute → save evidence → report → review → continue. Status queries remain useful for recovery, timeouts and final verification.
Here, distributed coordination means that execution and feedback live in the individual AI conversations. AgentRelay remains a shared local communication service, and the controller remains responsible for goals, dependencies and acceptance. There is no automatic shared context or infrastructure-free cluster.
- Controller assigns the goal
- Workers execute independently
- Completion reports arrive directly
- Controller reviews and continues
What the six-direction Windows test actually verified
On November 12, 2026, AgentRelay 2.11.0 connected existing Codex, Antigravity and DeepSeek Harness conversations in test TRILINK-20261112-84f099. All six directions received messages. Antigravity and Harness exchanged random challenges and returned matching values, then their A_DONE and H_DONE reports arrived inside the same active Codex turn.
This separates three kinds of evidence: accepted delivery, confirmed reception and a worker’s completion report. A successful HTTP response alone proves neither that the peer read the message nor that the business task is complete. The communication test did not establish software-task quality, a cost benchmark or unattended uptime.
Give every task a return address and a recovery rule
The local agentrelay-session-bridge skill defines a task ID, one writer per scope, completion criteria, replyProvider, replySessionId and an explicit idle-receipt policy. Reports include the conclusion, artifact locations, validation and unresolved issues. Download the public collaboration Skill below and install the whole folder. It includes the task contract, relay tools and simulated checks. The separate integration Skill connects an extension or app to AgentRelay.
Use POST /steer with strict:true for an active turn. An idle conversation needs the normal /{provider}/jobs entry to wake it. If a strict send explicitly returns 409 no_active_turn, a permitted reply can queue one result notification to the same conversation. A timeout or incomplete receipt is unknown: recover the original operation instead of blindly sending it again. Never acknowledge an acknowledgement.
Start with one small, verifiable handoff
Install the Windows x64 EXE, open and sign in to the supported clients, then choose existing project conversations. Start with a read-only review or another bounded task, specify where its evidence will be saved and how it must report back. Confirm that the controller actually receives the report before extending the workflow. Version 2.11.0 upgrades manually through the EXE; an EXE auto-update path is not promised.
Active-turn direct delivery and Antigravity’s idle-task wake-up were tested. An idle Codex controller automatically continuing after a queued completion report was not included in this test. Long unattended operation, high concurrency and crash recovery need separate validation. Keep clients and AgentRelay running, preserve receipts and let the controller decide when the goal is complete.
Install the Skill that connects your AI team
agentrelay-session-bridge lets a coding assistant delegate to existing Codex, Antigravity and DeepSeek Harness conversations and receive their completion reports. Use this for AI-to-AI coordination; agentrelay-integration is for connecting your own extension or app.
Requires Windows, local AgentRelay 2.11.0+, Node.js 22+, and supported AI clients running and signed in. Choose existing conversations. The scripts read the local AgentRelay token; never paste it into chat, website fields or public files.
Download and extract the whole folder
Keep agentrelay-session-bridge intact. This ZIP contains Skill source and scripts, not a portable AgentRelay application.
Install it in your assistant’s Skill directory
For Codex, place the whole folder under ~/.codex/skills/ so that ~/.codex/skills/agentrelay-session-bridge/SKILL.md exists. Other assistants use their own supported Skill directory. Reopen the assistant or refresh its Skills list if needed.
Invoke the Skill with one small goal
Use $agentrelay-session-bridge, describe the goal and choose existing conversations. Let the assistant check the local service and perform a bounded test. Downloading the ZIP alone does not establish a connection.
Give this instruction to your coding assistant
$agentrelay-session-bridge Check my local AgentRelay service and list existing Codex, Antigravity and DeepSeek Harness conversations. Help me choose one read-only review task. Read the controller’s current runtime identity and set replyProvider and replySessionId from it. Include the goal, scope, completion criteria and artifact location. Require the worker to save its result and proactively send DONE or BLOCKED to this controller. Use strict Steer for an active turn. If explicitly permitted, queue one result notification only after a definite no_active_turn response. Unknown delivery must recover the original operation without blind resend. Report actual reception and test results. Do not commit, publish or operate external accounts.
Task contract and completion command
Fill these fields from the current task
taskId: <unique task ID>
goal: <bounded goal>
scope: <read-only or explicitly owned files>
completionCriteria: <evidence to verify>
replyProvider: <codex | antigravity | dsh>
replySessionId: <current controller session from runtime>
receiptIdlePolicy: <wake-on-confirmed-idle only if authorized>
result: <UTF-8 file with DONE/BLOCKED, artifacts, checks and unresolved issues>New tasks and idle activation use /{provider}/jobs. Strict POST /steer targets an active turn. The optional --wake-on-idle below applies only to an explicitly authorized result notification after 409 no_active_turn; omit it otherwise. Accepted delivery, peer reception and completed work are separate checks. The idle-Codex continuation path was not covered by the six-direction test.
Worker: return the saved result to the controller
node "<skill-dir>/scripts/relay-direct.cjs" reply --provider <replyProvider> --session "<replySessionId>" --file "<UTF-8-result-file>" --wake-on-idle --operation-id "<unique-UUID>"Need to connect an extension or app? Use the integration Skill.
Official AgentRelay release notes · Open AgentRelay
Your AI clients can work as a team.
Build the loop first. Expand the team after one task returns with evidence.
