Signing you in...

Please wait while we verify your authentication

Article · Thursday, October 22, 2026

Agentic Coding

Top tech stories today across software, hardware, AI, and product launches. Senior engineer audience — skip rumour churn and pre-announcement leaks. Lead with shipping-now stories and what changed for builders.

By Marius BongartsTech138 editions
← See today's latest
Editions
4 / 138
Generated by AI overnight from public sources, refreshed daily.
Agentic Coding
Thursday, October 22, 2026
AI Agents - Agentic Coding

Multi-agent orchestration hits its stride with herdr

1 min read

Shape-Build-Review-Ship Loop

A year of agentic coding condensed into one workflow pattern.

Damian Galarza published his updated stack after 12 months running Claude Code, Codex, and Pi together through herdr—a terminal multiplexer built specifically for agent coordination [Source: Galarza]. His four-stage loop—shape, build, review, ship—puts the leverage where you'd expect: shaping work upfront with skills like grill-with-docs ensures shared understanding before code gets written, while verification-first building lets him run parallel agents safely. He uses specialist review agents checking architecture, testing, security, and docs before anything ships.

The pattern maps cleanly onto the workflows you've already been building.

Context Stacking Strategy

Your AGENTS.md file is just one layer of the context puzzle.

Galarza's setup shows how to stack context sources beyond the usual project-level instructions: Linear MCP pipes in tickets, Sentry CLI surfaces production errors, Fireflies adds meeting recordings, and Notion or Obsidian feeds team and personal knowledge [Source: Galarza]. He stores global agent instructions across all three tools—Claude Code at ~/.claude/CLAUDE.md, Codex at ~/.codex/AGENTS.md, Pi at ~/.pi/agent/AGENTS.md—keeping each tool's context scoped to its strengths.

Worth auditing what context your agents are actually seeing each session.

Tool-Specific Routing

Each agent tool earns its place by doing one thing well.

The workflow splits tools by task type rather than using one for everything: Claude Code handles frontend and design work where its visual reasoning shines, Codex runs the Sol model for efficient performance on straightforward tasks, and Pi orchestrates local models when you need more control [Source: Galarza]. The herdr multiplexer lets you coordinate all three without context-switching between terminals. He emphasizes verification over planning in the build phase—poteto-mode keeps parallel agents honest.

If you're still using one tool for everything, this routing split is your next experiment.

Sources
Agentic Coding Workflow: Claude Code, Codex, Pi - Damian Galarza
Agentic Coding Workflow: Claude Code, Codex, Pi - Damian Galarza
13 hours ago ... Each agent looks in a different place: Codex: ~/.codex/AGENTS.md; Claude Code: ~/.claude/CLAUDE.md; Pi: ...
damiangalarza.com
AI Summary

Damian Galarza updates his agentic coding workflow after 12 months of development, reflecting on how models like Opus 4.5, Fable 5, and GPT-5.6 have transformed agent capabilities. He manages a multi-tool stack including Claude Code, Codex, and Pi, each optimized for specific tasks—using Claude Code for frontend and design work, Codex for efficient model performance (Sol), and Pi for orchestrating local models through herdr, a terminal multiplexer built for agent coordination. Context remains critical for agent success. Galarza structures context through Linear MCP for tickets, Sentry CLI for production errors, Fireflies for meeting recordings, and Notion/Obsidian for team and personal knowledge bases. He stores global agent instructions in AGENTS.md files (~/.claude/CLAUDE.md for Claude Code, ~/.codex/AGENTS.md for Codex, ~/.pi/agent/AGENTS.md for Pi). His workflow follows a four-stage loop: shaping work with skills like grill-with-docs and domain-modeling to ensure shared understanding before coding; building with pstack (poteto-mode) for parallel agent orchestration emphasizing verification over planning; reviewing through specialist agents checking architecture, readability, testing, maintainability, security, and documentation; and shipping via conventional commits and GitHub stacked pull requests. He emphasizes that shaping and verification are the strongest leverage points—well-understood problems enable agents to work independently, and verification-first approaches enable safe delegation and orchestration across teams of agents.

Visit source
Compiled overnight by MorningMail.aiDelivered at 04:55 AM