Agent Harness Releases, September 2026: Claude Code, Codex, Gemini CLI, Hermes Agent and More

An agent harness is the software around a model that turns it into a working agent: the tool loop, memory, context management, permissions and approvals. Harness updates now move results as much as model updates do. Here is what shipped through September 26, 2026.

Short answer: September 2026 was a hardening month. The big coding CLIs shipped new models and a large volume of reliability and security fixes. The open harnesses (Hermes Agent, NVIDIA's SoL-Pi) focused on doing the same work with fewer tokens. Every version and date below was checked against the project's own release notes.

Release table

Claude Code 2.1.283

Anthropic's changelog for 2.1.283 is mostly aimed at teams running Claude Code at scale:

• Model governance. A new availableModelsMatch: "exact" managed setting blocks new model versions until an admin lists them, and deniedModels blocks specific models outright.
• Prompt audit. /doctor prompt-audit scans CLAUDE.md files, skills, agents and commands for prompting patterns written for older models.
• Gateway and telemetry. An opt-in prompt-ID header lets LLM gateways group the requests behind one user prompt, and OpenTelemetry can now include MCP, WebFetch and WebSearch outputs.
• MCP reliability. Fixes for lost progress notifications on background tools, stdio servers left running after exit, and a remote server becoming unusable after a brief 404 during a redeploy.

Codex CLI 0.157

OpenAI's Codex releases show 0.157.0 on September 25 and 0.157.1 on September 26, with 0.159 alphas already shipping:

• GPT-6 Sol and Luna models, including through Amazon Bedrock.
• Background server starts automatically for interactive sessions, with recovery options.
• Fork a conversation with the f shortcut while keeping your draft; fullscreen transcripts are now the default.
• Fixes for voice sessions, file-upload retries (5-minute timeouts) and proxy routing for realtime and web search.

Gemini CLI 0.61 and 0.62 preview

Google's Gemini CLI releases put security first. v0.61.0 (September 23) prevents indirect prompt injection through build-file changes and hardens filesystem boundaries. v0.62.0-preview.0 adds Gemini 3.8 Flash and 3.5 Flash Lite, structured MCP tool-call signatures, and better OAuth refresh-token handling.

Hermes Agent v0.21.5

Nous Research's MIT-licensed Hermes Agent shipped four patch releases in September (v0.21.2 to v0.21.5) after the v0.21.0 "Pantheon" release on August 31. Pantheon added multi-agent Bot Mode, agent-to-agent messaging with hermes peer, scheduled jobs with persistent memory, and desktop browser automation. The patches fixed state.db reliability and isolation between profiles. See our Hermes and OpenCode swarm guide for deployment patterns.

NVIDIA SoL-Pi

SoL-Pi is an MIT-licensed extension for the Pi coding agent from NVIDIA, NTU and MIT researchers. It adds four opt-in mechanisms that auto-research loops discovered: Action Fusion, ObservationPack, an Evidence-Preserving Reducer and Online Context Compact. MarkTechPost reports 44.7–49% less token traffic and about 33% lower API cost, while keeping about 94% of base Pi's score.

Microsoft Agent Framework Harness

Released on July 22, the Agent Framework Harness wraps any chat client with function calling, history persistence, context management, to-do planning, file memory, skills discovery, web search, tool approval and OpenTelemetry. It's available in Python and .NET.

What this means

• The harness is now a competitive lever. OpenAI's Codex platform post reports harness changes alone taking GPT-5.6 Sol from 13.3% to 38.3% on ARC-AGI-3.
• Governance is catching up. Model pinning, deny lists, prompt-injection hardening and audit telemetry are what companies need before rolling agents out widely.
• Token efficiency is the next frontier. Coding agents now run for hours, and harness-level savings add up fast.

For the open-source side of the month, read our open-source AI agent roundup. To compare harnesses side by side, see the agent harness matrix.

Frequently asked questions

What is an agent harness?

An agent harness is the runtime that wraps a language model and makes it an agent. It handles tool calls, conversation history, context-window management, planning, memory, permissions and human approval. Claude Code, Codex CLI, Gemini CLI and Hermes Agent are all harnesses.

What is the latest Claude Code version?

As of September 26, 2026, the latest Claude Code release in Anthropic's public changelog is 2.1.283. It adds managed settings to pin or deny specific models, a prompt-audit mode in /doctor, and many MCP reliability fixes.

What is new in Codex CLI 0.157?

Codex CLI 0.157.0, released September 25, 2026, adds the GPT-6 Sol and Luna models with Amazon Bedrock support, makes fullscreen transcripts the default, starts a background server automatically for interactive sessions, and adds a shortcut to fork a conversation. 0.157.1 followed on September 26.

Does the harness matter more than the model?

It increasingly matters as much. OpenAI reported that harness changes alone (retained reasoning and context compaction) raised GPT-5.6 Sol's ARC-AGI-3 score from 13.3% to 38.3% while cutting output tokens sixfold.

Which agent harness should I use?

It depends on the job. Coding in a terminal points to Claude Code, Codex CLI or Gemini CLI. A self-hosted, always-on assistant with messaging and scheduled jobs points to Hermes Agent. Embedding an agent in your own product points to the open-source Codex app-server or the Microsoft Agent Framework Harness.