Agent Harnesses logopeterindia.net
Knowledge Hub

Agent Harnesses

A directory of 12 entries: the five memory-generation harnesses on the FrontierMemory leaderboard, the coding harnesses paired with them, and the benchmark's own pages. Links go to each vendor's site.

19 topics•3 categories•Updated October 2026

The Five FrontierMemory Harnesses

5 entries

The memory-generation harnesses ranked on the FrontierMemory leaderboard, best first. Scores and task costs are Letta's, from frontiermemory.com.

01Letta DreamBackground dreaming subagents review recent conversations and consolidate lessons into memory, which Letta says lives in MemFS, a git-backed memory filesystem the agent can inspect and edit; it can run after a set number of agent steps or when the context window is compacted. FrontierMemory ranks it first at 73.3% and $2.80 average task cost. Letta runs the benchmark. 02Prime AgentPrime Intellect's open-source (MIT) agent for coding and long-running tasks, built around a recursive-language-model design with a persistent Python REPL. Its /refine command reviews the current trajectory and applies small updates to supplemental harness state, with snapshots for rollback. FrontierMemory ranks it second at 60.0% and $3.74. Its README warns that it is not a security sandbox, and GitHub lists a nightly build (v0.10.2-beta.3, 22 October 2026) as the newest release. 03OpenAI Agents SDKOpenAI's open-source (MIT) Python framework. Its sandbox-agent memory lets later runs learn from earlier ones: a two-phase process summarises conversations and consolidates them into MEMORY.md and memory_summary.md in the sandbox workspace. The docs page does not use the word dreaming, though FrontierMemory describes it that way. Ranked third at 40.0% and $0.82. v0.24.0 was released on 15 October 2026, according to GitHub. 04Claude Managed AgentsAnthropic's hosted agent service. Dreaming, announced on 19 May 2026, is a scheduled process that reviews agent sessions and memory stores for patterns such as recurring mistakes and reorganises memory; Anthropic lists it as a research preview with access by request. Ranked fourth at 33.3% and $3.21. 05Hermes AgentNous Research's open-source (MIT) agent that creates skills from complex tasks and improves them in use, with agent-curated memory, session search and user modelling across sessions. FrontierMemory uses its built-in learn feature and ranks it fifth at 26.7% and $2.79. v0.22.5 was released on 23 October 2026, according to GitHub.

Task Harnesses Paired With Them

3 entries

Coding agents that FrontierMemory says it pairs with the memory harnesses to attempt the downstream tasks.

06Letta CodeLetta's open-source (Apache-2.0) stateful coding agent, which Letta says it uses for task evaluation, paired with Letta Dream. GitHub lists v0.34.23, from 23 October 2026. 07OpenAI CodexOpenAI's coding agent, which FrontierMemory pairs with the OpenAI Agents SDK memory feature. 08Claude CodeAnthropic's coding agent, which FrontierMemory pairs with Claude Managed Agents Dreaming.

About the FrontierMemory Benchmark

4 entries

Letta's benchmark for memory quality, plus its task list, write-up and submission page.

09FrontierMemoryLetta's benchmark site, with model and harness leaderboards. It says agents are given past trajectories, generate memory, and are scored on how much that memory helps with downstream tasks. 10FrontierMemory 1.0 write-upLetta Research's October 2026 article with per-model scores for each harness. It says each harness was evaluated with matched memory and task models at max reasoning effort, with one attempt per task across 15 tasks. 11FrontierMemory tasksThe benchmark's task list. Letta says it draws coding and knowledge-work tasks from public benchmarks, including HarborIndex, DeepSWE and SkillsBench. 12Submit a harness or modelLetta says new memory-generation harnesses can be added through an adapter pull request.

Related Pages

7 entries

Related directories on this site.

13A Meta-Harness for AI AgentsHarnesses as swappable parts, with Omnigent as the example. 14AI Coding AgentsCoding agents such as Claude Code and Codex. 15AI Agent MemoryMemory systems for AI agents. 16AI Agent FrameworksFrameworks for building agents. 17Enabling Tools of AI Coding AgentsTools that coding agents rely on. 18Open-Source Benchmarks for AI Coding AgentsBenchmarks for coding agents. 19Agentic AIAgentic AI hub.

Frequently Asked Questions

What is an agent harness?

An agent harness is the code wrapped around a model that turns it into an agent: the loop, tools, context handling, memory and permissions. Claude Code, Codex, Letta Code, Hermes Agent and Prime Agent are examples. The same model can behave very differently in different harnesses.

Which five harnesses does FrontierMemory rank?

Letta's FrontierMemory leaderboard ranks Letta Dream (73.3%), Prime Agent (60.0%), the OpenAI Agents SDK (40.0%), Claude Managed Agents (33.3%) and Hermes Agent (26.7%). These are memory-generation harnesses, meaning they turn past agent trajectories into memory. Scores and task costs are as published by Letta.

Are these the only agent harnesses?

No. They are the five harnesses on one benchmark's leaderboard, chosen for how they generate memory. Coding harnesses such as Claude Code, Codex and Letta Code do the downstream tasks, and many other harnesses exist. See the AI Coding Agents, AI Agent Frameworks and Meta-Harness pages on this site.

How should I read the FrontierMemory results?

With care. Letta created the benchmark, its own Letta Dream ranks first, and Letta says it is used by default for memory generation of submitted models. The write-up says each harness had one attempt per task across 15 tasks, so differences can be noisy. The OpenAI Agents SDK and Claude Managed Agents were each scored with one task model, while Letta Dream, Prime Agent and Hermes Agent were scored with two, and the ranking uses the higher result.

What is dreaming in an agent?

Dreaming is background processing that reviews past agent sessions and consolidates useful lessons into memory, so later runs start with better context. Letta, Anthropic and, according to FrontierMemory, OpenAI's Agents SDK memory all offer a version, but each stores and reviews memory differently.

How current is this directory?

Links were checked on 23 October 2026. The FrontierMemory scores come from frontiermemory.com and Letta's write-up on that date, and release information comes from GitHub. Products change quickly, so confirm details on each project's own site.