2026 harness layer for programming agents: loop, tools, memory, sandbox, evals, and orchestration — who ships it, what the evidence shows, what is contested.

Best agent harnesses for programming (2026)

Generated by Grok Bot research on 2026-08-27. WebSearch + WebFetch + native X. Treat as raw material — review before promoting.

Summary

A programming agent harness in 2026 is the software around a model — loop, tools, context/memory, sandbox, hooks, and orchestration — not the model itself. Anthropic and LangChain both use that split. The leading shipped products are Anthropic’s Claude Code / Claude Agent SDK, OpenAI’s Codex CLI plus AGENTS.md, and the MIT, model-agnostic OpenHands Software Agent SDK. Official Terminal-Bench 2.0 shows the same model swinging ~16–18 points across harnesses; Codex CLI + GPT-5.5 is the highest official vendor CLI (82.2%), while Claude Code + Opus 4.6 sits at 58.0%. Contested: CLAUDE.md vs AGENTS.md, whether to add or strip scaffolding, and whether eval-optimized harnesses transfer to daily coding.

Findings

What a programming agent harness is

Anthropic (2 Apr 2026) defines an agent harness as “the software scaffolding around a model: the loop, tools, context management, and guardrails that turn raw intelligence into a working agent,” and harness design as deciding what belongs in that scaffolding and what to remove as models improve. LangChain (10 Mar 2026) states the same boundary as “Agent = Model + Harness” and “if you’re not the model, you’re the harness,” listing system prompts, tools/skills/MCP, filesystem/sandbox/browser, subagent orchestration, and hooks (compaction, continuation, lint).

That is distinct from an IDE recap. The product (Claude Code, Codex, Cursor) is one packaging of a harness. The reusable substrate is now also sold as an SDK: Anthropic’s Claude Agent SDK (“the same tools, agent loop, and context management that power Claude Code”), OpenHands’ Software Agent SDK (Python/REST engine behind OpenHands CLI and Cloud), and OpenAI’s Codex runtime with permission profiles and AGENTS.md.

Who ships the leading ones

Claude Code / Claude Agent SDK (Anthropic). Official docs list built-in file/shell/web tools, hooks, subagents, MCP, permissions, sessions, skills/commands/memory from .claude/ and ~/.claude/, and plugins (Agent SDK overview). Dynamic workflows (2 Jun 2026) let Claude write a JavaScript orchestrator that spawns isolated subagents — classify-and-act, fan-out-and-synthesize, adversarial verification, tournament, loop-until-done — to fight agentic laziness, self-preferential bias, and goal drift. Trigger word ultracode. Authors warn workflows use more tokens and are for high-value tasks, not every edit.

Codex CLI (OpenAI). Official permissions treat sandbox as a first-class policy language: built-ins :read-only, :workspace, :danger-full-access; filesystem read/write/deny; optional network proxy + domain allowlists; macOS Seatbelt, Linux/WSL bubblewrap+seccomp, Windows elevated sandbox users. Profiles do not compose with older sandbox_mode. Official AGENTS.md is the instruction chain (global ~/.codex, then repo-root-to-cwd walk, override files, 32 KiB default cap, fallback filenames). On Terminal-Bench 2.0, Codex CLI + GPT-5.5 is rank 4 at 82.2%±2.2 (23 Apr 2026).

OpenHands Software Agent SDK (All Hands / OpenHands). Official SDK is MIT, model-agnostic, coding-specific, with bash/file/browser/MCP tools, Docker/K8s server, and an OpenAI-compatible endpoint. Architecture splits openhands.sdk (agent loop, tools, skills, condenser, security), openhands.tools, openhands.workspace (local / Docker / remote), and openhands.agent_server. Same agent code swaps LocalWorkspace → DockerWorkspace → RemoteAPIWorkspace. The product GitHub page reports 85,281 stars (thin fetch; README body not extracted). Docs claim OpenHands is “becoming the preferred harness for evaluating LLMs on coding tasks” (SDK); Terminal-Bench 2.0 places OpenHands + Claude Opus 4.5 at 51.9%±2.9 (leaderboard) — strong open platform, not the TB winner.

Eval-specialized / research harnesses. Terminal-Bench 2.0 is itself a Harbor eval harness (harbor run … -k 5, no timeout/resource overrides). Top rows as of the fetched page: NexAU-AHE + GPT-5.5 84.7%; LemonHarness 84.5%; Capy + GPT-5.5 83.1%; then Codex CLI. Also listed: Factory Droid, JetBrains Junie CLI, LangChain Deep Agents (66.5% with GPT-5.2-Codex), Gemini CLI, Warp, Goose (Block), OpenCode, Princeton Mini-SWE-Agent (42.5% with Sonnet 4.5). These are harnesses optimized for containerized terminal tasks, not necessarily daily IDE use.

Layer map (loop, tools, memory, sandbox, orchestration, evals)

Loop. Anthropic and OpenHands both describe a ReAct-style plan → tool → observe → repeat. OpenHands names this the reasoning-action loop inside Agent + Conversation. LangChain adds Ralph-style “don’t let the model exit” hooks that reinject the goal against filesystem state. Anthropic long-running apps (24 Mar 2026) contrasts context resets (clean slate + handoff file; needed for Sonnet 4.5 “context anxiety”) with compaction (summarize in place; enough for Opus 4.5+).

Tools. Anthropic’s load-bearing claim: Claude 3.5 Sonnet hit 49% SWE-bench Verified with only bash + a text editor, and Claude Code is grounded in those same tools (harness design). Skills, programmatic tool calling, and the memory tool are compositions of bash/editor. Dedicated typed tools stay for UX, observability, or hard-to-reverse actions. Codex encodes the same idea as schema-first exec plus a separate permission engine (permissions).

Memory / instructions. Two competing file standards. Codex official: AGENTS.md with global + nested overrides. Claude Code official: CLAUDE.md / .claude/ skills (Agent SDK; dynamic workflows mine sessions back into CLAUDE.md). Anthropic reports a memory-folder lift on BrowseComp-Plus from 60.4% to 67.2% for Sonnet 4.5, and compaction scaling Opus 4.6 to 84% on BrowseComp vs Sonnet 4.5 stuck at 43% (harness design). Discourse: Shopify CEO Tobi Lütke threatened to ban Claude Code until it reads AGENTS.md (X, primary); Claude Code staff Thariq Shihipar said they will add AGENTS.md/system-prompt hackability, but argued model families are not interchangeable so Claude.md stays default (X, primary).

Sandbox. Codex’s documented default is least-privilege local execution with platform enforcement that refuses a command if the policy cannot be applied (permissions). OpenHands’ production path is container/remote isolation (SDK architecture). Claude Code’s published security story is permissions + auto-mode (a second Claude judging bash) rather than a full OS jail in the fetched docs (harness design). OpenAI’s official X post on the Hugging Face incident is a primary that existing agent safeguards can fail in the wild (X, primary).

Orchestration. Two official Anthropic patterns: (1) strip the harness and let the model orchestrate via code exec — BrowseComp 45.3% → 61.6% when Opus 4.6 filtered its own tool outputs (harness design); (2) add a planner/generator/evaluator multi-agent harness for long-horizon apps, with sprint contracts and Playwright QA (long-running apps). Dynamic workflows are the productized version of (2) (dynamic workflows).

Evals. Harbor / Terminal-Bench 2.0 is the public harness-vs-harness table. Same-model swing on the fetched page: Claude Opus 4.6 is 58.0% in Claude Code, 74.7% in Terminus-KIRA, 75.3% in Capy, 76.4% in Stanford IRIS Meta-Harness. GPT-5.5 is 82.2% in Codex CLI vs 84.7% in NexAU-AHE. LangChain claims a harness-only jump of Deep Agents from top-30 to top-5 on TB 2.0 (anatomy post); the fetched leaderboard has Deep Agents + GPT-5.2-Codex at 66.5% (rank 27).

Cost and “best” are not the same axis

Anthropic’s own long-running experiment: solo Opus 4.5 built a broken game maker in 20 minutes for $9; the full planner/generator/evaluator harness ran 6 hours for $200 and produced a playable app (long-running apps). The simplified Opus 4.6 DAW harness was ~4 hours / $124.70. Discourse (not a lab measurement): Xiaoyin Qu claimed the same task cost $150 on Claude Code vs $2 on DeepSeek’s harness and launched AgentSky as an “OpenRouter for agents” (X, discourse).

Claude Code creator Boris Cherny’s recent primary posts are about using the loop, not ranking harnesses: shared memory got simpler (X, primary); Opus is used to “hill climb” profiler-closed loops until a numeric target (X, primary).

Contradictions and open questions

Vendor default vs eval-optimized harness. TB 2.0 says Claude Code is not the best container for Opus 4.6 on that bench; specialized harnesses beat it by ~16–18 points. Anthropic and LangChain both say models are post-trained inside a vendor harness and can overfit (LangChain cites Codex apply_patch). Unresolved: how much of the TB gap is reward-hacking / timeout policy vs a real daily-coding gap. Harbor forbids timeout/resource overrides; that does not prove transfer to a laptop repo.
Add scaffolding vs strip it. Same lab, same month-range: long-running apps shows a three-agent harness 20× more expensive and qualitatively better than solo; harness design says prune dead weight (“what can I stop doing?”) because context-reset scaffolding for Sonnet 4.5 became useless on Opus 4.5. Dynamic workflows then adds orchestration. The consistent rule is “complexity must earn its keep,” not a universal more-or-less.
CLAUDE.md vs AGENTS.md. OpenAI documents AGENTS.md as the instruction standard (official guide). Anthropic still ships CLAUDE.md as native. Tobi (primary) vs Thariq (primary) vs Kun Chen (discourse: “industry already has a standard”) (Tobi, Thariq, Kun Chen). Codex already has project_doc_fallback_filenames; Claude Code’s promised AGENTS.md support was “working on it” as of 25 Aug 2026, not fetched as shipped.
OpenHands “preferred eval harness” vs the public TB table. OpenHands SDK docs claim preferred-eval status and SOTA on SWE-bench/SWT-bench/multi-SWE-bench. TB 2.0 has OpenHands well below Codex CLI and several closed CLIs. Those claims can both be true if they refer to different benches and dates; this pass did not fetch SWE-bench.org numbers.
Karpathy. Recent @karpathy posts in this window are not a harness ranking. Closest discourse: tearing down abstractions now that agents can do the math and verification (X). No 2026 Karpathy “best harness” primary found.
Gaps this pass did not close. OpenAI Codex product overview (developers.openai.com/codex) timed out. Grokipedia search 502’d twice; tool-use and Model_Context_Protocol page-preview returned found: false. SWE-bench.org / OpenHands Index not fetched. Cursor / Copilot / Windsurf / Devin treated as already-covered product layer and not re-surveyed. Harbor’s own Terminus 2 is an eval agent, not a daily driver.

Provenance

Method: Grok Bot / WebSearch / WebFetch / native X Generated: 2026-08-27 Rounds: 3 (Round 1 definition + shippers; Round 2 official harness posts + TB 2.0 + Codex sandbox; Round 3 AGENTS.md + OpenHands architecture + X hydration). Early-exit not taken: Round 2 opened the CLAUDE.md/AGENTS.md and TB-swing questions that Round 3 resolved enough to write. URLs fetched: 11 successful, 4 failed (15 attempts). X: 7 cited permalinks (plus account scans that produced no extra harness claims).

Web sources:

Harnessing Claude's intelligence — Anthropic definition + prune-vs-add + BrowseComp numbers
The anatomy of an agent harness — Agent = Model + Harness
Claude Agent SDK overview — same loop as Claude Code
A harness for every task — ultracode dynamic workflows
Harness design for long-running apps — reset vs compaction; $9 vs $200
Codex permissions — sandbox policy language
AGENTS.md — instruction chain
OpenHands Software Agent SDK — MIT model-agnostic SDK
OpenHands SDK architecture — workspace swap
Terminal-Bench 2.0 leaderboard — harness swing
OpenHands GitHub — thin: stars only

X sources:

X post by @tobi (2026-08-25) — AGENTS.md ultimatum — primary
X post by @trq212 (2026-08-25) — Claude.md stays default — primary
X post by @bcherny (2026-08-26) — shared memory simpler — primary
X post by @bcherny (2026-08-23) — hill-climb loops — primary
X post by @OpenAI (2026-08-26) — HF incident safeguards — primary
X post by @karpathy (2026-08-21) — tear down abstractions — discourse
X post by @quxiaoyin (2026-08-24) — $150 vs $2 — discourse
X post by @kunchenguid (2026-08-25) — industry standard — discourse

Grokipedia: not used (search 502; page-preview found: false)