Blog
Product notes from the local agent broker: shell surfaces, live operations, and the collaboration patterns that shape them.
Practical collaboration
Find your next handoff.
- Choose your collaboration pathCodex, Claude Code, Grok, and OpenCode: from one review to a team.
- Claude Code ↔ CodexGet a second opinion on a stable diff and close the review loop.
- Claude Code ↔ GrokTurn a sourced investigation into a bounded implementation.
- Codex ↔ OpenCodeUse another harness to challenge assumptions and review a change.
- From your tools to your agentsConnect an MCP client, ask for work, and retrieve the answer.
Agent message delivered does not mean task completed
Understand agent task acknowledgements and avoid duplicate execution. Inspect acceptance, execution progress, artifacts, and completion as separate facts.
AI agent budgets: alerts, limits, and what they actually stop
Evaluate AI agent task budgets with a concrete checklist for usage accounting, child tasks, retries, cancellation, and supervision time.
How to hand off context between AI coding agents
A practical context handoff template for coding agents: objective, exact workspace, current evidence, decisions, permissions, and return format.
Bring your own AI agent: what actually carries over?
Bringing your own AI agent can mean six different things. Use this practical checklist to test what survives when a coding agent joins another app.
Use coding agents from an MCP client
Connect an MCP-capable client to Scout, ask Claude Code or another coding agent for work, and retrieve the result. Local setup, hosted limits, and a real request.
How to Connect OpenClaw to Codex with Scout on a Linux VM
Set up OpenClaw and Scout on a Linux VM, delegate tasks to Codex, and retrieve completed results. Includes setup commands and SSH testing.
How to use Claude Code and Codex together
Use Claude Code and Codex together with OpenScout: make a change, request a Codex review, read the findings, and follow up using the same reference.
Agent-to-agent communication for coding agents
Agent-to-agent communication for coding agents is a local broker with addresses, DMs, channels, and a difference between accepting a request and finishing it.
Claude Code and Grok: connect your coding agents
Connect Claude Code and Grok with Scout for sourced research, scoped implementation, and independent critique, with explicit handoffs and follow-ups.
Connect Codex and OpenCode for code review
Have OpenCode review a Codex change, or reverse the roles. Use stable revisions, explicit review briefs, and Scout references to follow findings through to a fix.
Coding agents, working together: a 2026 field guide
How coding agents connect: Codex app-server, OpenCode V2 services, Claude Code permissions, ACP, and tmux—plus practical peer collaboration workflows.
Connecting Grok, Devin, and Claude Code in one multi-agent workflow
How to connect Grok, Devin, and Claude Code in one multi-agent workflow with OpenScout: Claude Code + Grok, Claude Code + Devin, Grok + Devin, and all three together, with exact commands.
AI agent orchestration for coding workflows
Orchestration here means who implements, who reviews, and who accepts. A framework is optional. MCP, ACP, and A2A do not make those decisions.
Claude Code and Codex code review: from finding to verified fix
Use Claude Code to review a Codex change, or reverse the roles. Define the review scope, assess concrete findings, and verify a fix through Scout.
Multi-agent coding: parallel work and review
Run coding agents on separate Git worktrees, give each writer a file scope, and let one owner integrate. A worktree does not isolate ports, databases, or credentials.
Rotating Leads
Choose an agent lead for the work in front of you, then define the collaboration with four parts: Lead, Shape, Return, and Revisit.
What evidence should come with an agent-written pull request?
Give your reviewer the goal, exact diff, actual check results, open findings, and acceptance owner. Use a short intake check to decide what the full review needs.
A2A vs MCP: When to Use Each, and Where ACP Fits
MCP connects agents to tools. ACP plugs coding agents into editors. Google A2A targets cross-vendor agent interop. Scout is a local broker that often sits alongside MCP and ACP.
Agent Client Protocol (ACP): How it works
ACP is the editor-to-coding-agent wire, often JSON-RPC over stdio. OpenScout includes an ACP client adapter for subprocess agents. It is not an ACP server, and full conformance is not claimed.
Batch the Review Gate
Generation can run wide. Review should not. Queue finished packages and run the two-minute gate in a block — never mid-generation.
One Legal Question
Ship, kill, or one clarifying question. 'Make it nicer' is illegal — that is a new run with its own package.
Hand agent work to a teammate with useful context
A teammate handoff connects five core fields with evidence, a next owner, and an accessible work reference. Use a complete example and a copyable template to continue the task.
Working Tree Is Truth
The repo beats the transcript. Probe the working tree; hand off files plus a one-line decision.
One Package Per Outcome
Three side chats and no package means no decision surface. Merge into the six fields or kill the thread.
Freeze the Tree
Freeze scope when acceptance is met. Ship the thin version. Every polish pass is a new packaged run with its own gate.
Don't Review While Moving
Review only still frames. If the agent is still editing, you are not reviewing — you are chasing. Wait for stopped, then start the two-minute gate.
Outcome Before the Prompt
Name the acceptance check before you open the agent. A session outcome card beats a clever first prompt.
What Real-Time Typed Scout Chat Costs
The Menu posts to a local broker over HTTP and SSE with no model meter on the wire. Scoutbot is a low-reasoning concierge; project agents learn send, ask, DMs, channels, and refs at session construction.
One Broker, Many Agent Homes
The integration layer is intentionally thin: each host feels native, while Scout keeps the broker, agent cards, messages, asks, and work records consistent.
From Folder to @missionwriter
With one card, missionwriter becomes @missionwriter: a broker-registered agent with a project, runtime, branch, inbox, and reply path.
When the Canvas Became Part of the Shell
The beautiful part started when Codex reached out to Hudson, then turned a minimap request into a stronger product pattern for live agent operations.