Cloneable harnesses, WASM sandboxes and control planes that own the agent loop

GreyhavenHQ

Container-free deny-by-default sandbox for coding agents that enforces kernel filesystem, network and syscall isolation on Linux and macOS.

Use caseUse it to wrap Claude Code or similar CLIs so they cannot touch secrets outside the working tree.

cordum-io

Agent control plane and action firewall that enforces policy and human approval before risky tool calls, shell commands and production changes with audit evidence.

Use caseUse it to insert a deterministic governance layer in front of any framework before agents act.

rohitg00

Persistent memory layer for coding agents that ranks high on real-world session and temporal retrieval benchmarks.

Use caseUse it to give any MCP or CLI coding agent durable cross-session recall without pasting full history.

cordum-io

Open wire protocol for agent clusters that adds jobs, heartbeats, safety hooks and workflows over a message bus.

Use caseUse it to standardize how distributed agents coordinate and apply policy before dispatch.

zenml-io

Record-replay and improvement system that treats agent runs as evidence-rich sessions with evaluators and cohorts.

Use caseUse it to capture production agent traces, define evaluators and run bounded replay experiments.

mngaonkar

Tiny skill-based local agent harness that makes the model, tools and disk skills fully visible in one sitting.

Use caseUse it to learn and extend a minimal laptop agent loop with Anthropic-style skills folders.

GitHub

OpenAI’s GPT-6.1 Terra is generally available across Copilot surfaces for everyday agentic coding and terminal work at lower cost than the Sol variant. Early tests show stronger results on terminal and C# tasks than prior Terra models.

Shipped

Tejas Kumar

The essay argues that after the agent loop works, harness engineering reduces to four ownership questions: who decides done, who holds keys, who says stop, and who is blamed. It shows how moving those decisions into deterministic harness code separates model bugs from scaffolding bugs.

Read