26 October 2026
Cracks in the Agent Loop: predictable, auditable, recoverable agents
A 100-call run changed 1,951 lines and opened zero pull requests. The fix was not a bigger iteration limit — it was a tighter loop and a smaller blast radius: contracts, budgets, checkpoints, telemetry, and no unnecessary credentials. Lessons for anyone running agents.
agentsreliabilitysecurityobservabilityCI
21 October 2026
Mentat 2.0: plan, execute, review — an agent-driven development loop
A three-phase agent loop: n8n triggers one-shot containers that plan, implement, and review real pull requests — with CI as the gate and separate bot identities for development, review, and orchestration.
n8nDockerGoGitHub Apps
6 October 2026
Approving agent work from Slack
How mentat’s opencode plugin turns Slack into a remote control for long-running agent sessions — approvals, questions, prompts, and a loader reaction that tells you when it’s still thinking.
opencodeSlackSocket ModeTypeScript