Running company ops on 16 Claude agents: 3 patterns to prevent chaos
Salty-Set-5853 · reddit · 2026-08-26
The author shares practical experience using 16 autonomous Claude agents to manage daily operations (sales research, cost audits, comms drafting) for a small company, with the architecture open-sourced.
The Challenge: By week three, agents started contradicting each other, with one making a 3 AM decision based on information the author had reversed the previous afternoon.
Three Patterns to Fix Chaos:
- One Shared Log: Every agent appends to a single log file (activitylog.) after each run and reads it before acting. No direct agent-to-agent messaging. Crucially, human interventions (pricing changes, vendor drops) are also logged to keep the agents' world model current.
- OBSERVE Mode for External Actions: Agents touching money, external comms, or public channels do not act. They research, evaluate, draft, and raise an escalation for human approval. Gating the write path allows higher autonomy elsewhere and prevents costly actions based on wrong premises.
- Maintainer Agent Constraint: A weekly maintainer agent reviews logs to propose rule updates but cannot promote them itself. This prevents an infinite loop of self-improvement that leads to decision drift and wasted costs.
Advice: Start with two agents that genuinely need coordination and nail the logging mechanism first.
More from coding & agent
- Best practices for Agent skills: personal assistants may require dozens of skills — intellectronica · 2026-08-26
- Livestream recap: Stripe Projects, Netlify CLI deploys, and Claude Code workflows — thisiskp_ · 2026-08-26
- YC-Backed Prized Runs Cloud-Based Coding Agents 24/7 — ycombinator · 2026-08-26
- Anthropic Cuts 80% of Claude Code's Prompt: Models Outgrew Scaffolding — ialijr · 2026-08-26
- AntiGravity update adds right pane with multiple terminals — rseroter · 2026-08-26
- Anthropic to make Claude Code more hackable with Agents.MD and system prompt edits — trq212 · 2026-08-26