Open-source agent harnesses address several different engineering problems: running a general-purpose agent, embedding an agent in an application, coordinating specialist agents, and adapting training environments. AWS Strands, TrueFoundry’s TrueForge, EverMind AI’s Raven, and EnvHarness illustrate those differences. Their architectures offer a more useful starting point for selection than a single ranking of “best” harnesses.
For software and platform engineers, the central question is which part of the system needs structure. Strands packages an agent ready for customization; TrueForge separates runtime, API, and interface; Raven coordinates agents and proposes harness improvements; EnvHarness changes the environment around an agent. The available performance figures are vendor reports, with different workloads and incomplete evaluation details. They do not establish a winner across these projects.
Strands: a packaged general-purpose agent
AWS describes Strands as a general-purpose agent. Built on its Strands SDK, the harness is available through Python and TypeScript libraries and a CLI, with local use or deployment against a preferred AI provider. The Register’s coverage presents it as an assembled starting point rather than a collection of components developers must first wire together.
The runtime includes shell, file, and web tools. It automatically loads supplied skills, supports memory across runs, and can resume conversations using a session ID. A built-in helper agent handles open-ended subtasks, while a checklist tracks multi-step workflows, according to DevOps.com.
Its context controls are particularly concrete. The Register reports defaults that truncate tool results over 1,500 tokens and compact the context window when it surpasses 85 percent, with attempts to recover from context overflow. DevOps.com also describes offloading bulky tool results to files and reusable caches. These mechanisms show where AWS is intervening in the agent loop: the volume of information retained and passed back to the model.
That makes Strands worth examining when the immediate need is a customizable agent with tools, session continuity, and context handling already assembled. Those features describe the runtime’s behavior; they do not establish how reliably it will complete an organization’s particular tasks.
TrueForge: runtime, API, and embeddable interface
TrueFoundry describes TrueForge as the MIT-licensed runtime it open-sourced in August 2026 and uses behind its AskTFY agent. Its packaging emphasizes application integration through three pieces:
| Component | Documented role |
|---|---|
| Core server | Agent loop, streaming, approval gates for sensitive actions, subagent delegation, compaction, and sessions that survive reconnects |
| HTTP API and <code>@truefoundry/trueforge-sdk</code> | TypeScript access to the capabilities exposed through the UI |
| Chat UI and <code>@truefoundry/trueforge-ui</code> | Interface that can be used directly, themed, or embedded |
TrueForge treats a sandbox as a tool, creating one when the agent needs to execute code. It does not wrap the entire session in a container, according to TrueFoundry. That is a meaningful architectural distinction for teams deciding how execution should fit into a broader application.
The same vendor comparison describes OpenHands, formerly OpenDevin, as using an event stream: the agent reasons, emits an action, the environment executes it, and an observation returns. Each session receives its own Docker sandbox. The practical comparison is therefore between different execution boundaries. The supplied evidence does not demonstrate that either boundary produces superior security, reliability, or task success.
TrueFoundry also distinguishes the open-source TrueForge runtime from its Agent Harness offering, which combines that runtime with an AI Gateway providing access controls, budgets, guardrails, and unified traces. Engineers should keep that distinction in view when assessing which operational features are documented for the runtime and which belong to the combined offering.
Raven: coordination and proposed self-improvement
EverMind AI’s Raven repository describes an Apache License 2.0 project in pre-alpha, built to coordinate agents and support recursive self-improvement. Raven generates DAGs for complex tasks, manages dependencies and parallel execution, and turns multi-step collaboration into reusable workflows.
Its built-in specialists are Raven-Research, Raven-Code, Raven-Design, and Raven-Oncall, covering research, coding, visual design, and unattended workflow automation. It also documents connections through ACP, CLI, or OpenAI-compatible APIs, with presets for 13 third-party agents. The scope extends beyond supplying one agent loop: Raven aims to organize work across agents with different roles and interfaces.
EverOS supplies memory and context across sessions, including user context, agent experience, and world knowledge. SkillForge retrieves skills from local libraries, EverOS memory, and SkillHub. Proactivity combines event monitoring with scheduled execution for reminders and follow-up work. These are repository-described capabilities, rather than independently established outcomes.
Raven’s Evolver diagnoses failures, tests candidate changes, and retains improvements that outperform a baseline in reproducible evaluations, according to the repository. Its pre-alpha status matters when interpreting that ambition. The evidence does not provide measured improvement rates or establish sustained autonomous improvement.
A separate research project, RRSI, describes a more specific candidate-admission procedure. Its unified codebase evaluates two candidates per round on the full evolve set, recalibrates δ from an Opus baseline with z = 2, and admits candidates inside the noise band only for token savings or a new structural component. This provides a concrete example of rules governing proposed harness changes. The evidence does not establish that Raven implements RRSI’s procedure.
EnvHarness: changing the training environment
EnvHarness targets another layer entirely. VentureBeat reports that researchers from Google Cloud AI Research and academic partners developed the Apache 2.0-licensed framework to adapt training environments to an agent’s weaknesses.
Its programmable wrapper can change where the agent starts, what it observes, which actions it can take, and how long a task lasts. The underlying environment and its verifier remain intact. That makes environment adaptation its defining capability, rather than application packaging or coordination between specialist agents.
For engineers considering training workflows, the important distinction is what gets modified. EnvHarness changes the conditions presented to an agent; Raven’s Evolver describes changes to the harness itself. The supplied evidence offers no common benchmark for comparing their effectiveness.
What the performance reports actually establish
AWS claims Strands matches rival harnesses while using around a quarter fewer tokens, crediting default prompt caching and context management. DevOps.com reports Marc Brooker’s figure of up to 28% fewer tokens. Those formulations differ in precision and scope: the latter is a maximum, not an established average saving.
The Register names Claude Code, Codex, oh-my-pi, OpenCode, and DeepSeek Harness as comparators, noting their coding-agent positioning despite AWS’s general-purpose framing. It also reports Amazon’s plan to publish further benchmark details. The supplied evidence does not include that promised methodology, so it cannot establish how the result transfers to other tasks.
TrueFoundry reports a separate comparison on DevRev’s Enterprise-Bench: 14 cross-system tasks across three MCP servers, fresh sessions, and a blind LLM judge.
| Configuration | Reported cost per run | Reported token count |
|---|---|---|
| TrueForge with Opus 4.8 | $8.5 | 3.8M |
| Claude Managed Agents | $11.8 | 10M |
| TrueForge with GLM-5.2 | $2.9 | Not supplied |
TrueFoundry says the Opus 4.8 configuration solved the same tasks as Claude Managed Agents. It also reports using under a quarter of deepagents’ tokens on the same model. These are vendor-reported results; the excerpt does not clarify the token-count aggregation or supply a complete task-level breakdown. Changing the model introduces another variable, so the GLM-5.2 cost cannot isolate the harness’s contribution. Nor can these enterprise-task results be combined with AWS’s report into a reliable ranking.
Choosing the layer—and checking the output
TrueFoundry’s comparison also describes opencode as an MIT-licensed, model-agnostic coding agent with terminal, desktop, and IDE interfaces; Pi as an MIT-licensed harness from Earendil Works with Read, Write, Edit, and Bash tools; and deepagents as LangChain’s MIT-licensed harness built on LangGraph. These descriptions come from a vendor comparing its own product with alternatives, rather than independent testing.
A sensible shortlist follows the architectural need: Strands for an assembled general-purpose runtime, TrueForge for application integration, Raven for coordination experiments, and EnvHarness for adaptive training environments. Coding-focused projects warrant consideration when repository work is the central task.
Selection still leaves output validation to address. Matthew Boston’s engineering essay explains how linters catch mistakes and formatters reduce distracting diff noise, while emphasizing that a successful build says nothing about behavior. That distinction carries across the comparison: documented orchestration, context controls, and lower reported token use are useful properties, but task-specific evidence is still needed to judge whether an agent’s work is correct.
