Directory
30 curated milestones in harness engineering — launches, funding, protocols and open-source moves. Curated snapshot of Oct 2, 2026. GitHub star counts were refreshed from the GitHub API on that date; valuations and funding figures are approximate public reporting and should be re-verified before relying on them.
30 entries
DeepSeek releases a macOS and Windows Desktop app for DeepSeek Harness, its open-source agent harness where everything — tools, sub-agents, even the loop — is a plugin. The repo passed 240k stars within three months of its August launch, making it the fastest-growing harness in the field's history.
The Best of Agent Harnesses list ranks 167 agent harnesses weekly by capability, autonomy and recovery, ships an MCP server so agents can pick their own harness, and maintains a graveyard that flagged a 268k-star repo for suspected star manipulation. Curation itself becomes harness infrastructure.
Restate raises $20M to 'define the infrastructure layer for AI agents and workflows' — its lightweight durable runtime journals every agent step so invocations survive crashes. The round underlines durable execution as the best-funded sub-category of agent infrastructure.
EverMind AI releases Raven, a trusted, persistent, self-evolving multi-agent ecosystem built for recursive self-improvement — agents that modify their own harness under governance. The Show HN thread became a running debate on how much autonomy a harness should hand an agent over itself.
'Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents' surveys the field — loops, tool interfaces, context management, memory and recovery — giving the emerging discipline its first comprehensive academic map.
Arena.ai's HarnessTax study tests 21 model–harness combinations across Claude Code, Codex CLI and Pi, finding harness choice can significantly change cost even at similar task success rates. Weeks earlier, SWE-bench Pro analysis showed swapping harnesses moved pass@1 more than model upgrades (23%→52% on GLM-5.2) — and that harness rankings barely transfer across models.
PrimeIntellect open-sources Prime Agent, a recursive-language-model agent whose harness runs sub-agents as function calls inside a persistent IPython REPL. The launch post reports its harness taking Opus 5 from ~30% to 95.5% on ARC-AGI-3 — the sharpest evidence yet that harness engineering can beat model selection.
a16z leads a $20M seed into Runta at a $100M+ valuation for its agent execution layer — sandboxes, spending caps and audit trails that treat agents like children who need supervision. Martin Casado called it 'core systems software rebuilt for agents'; the round came after prompt-injection containment failures made agent governance urgent.
xAI open-sources Grok Build — a Rust coding agent harness and fullscreen TUI around Grok models — unusually early for a first-party tool. It joins Claude Code, Codex and Gemini CLI as the fourth major first-party harness, now with a fourth open one.
Arcade.dev raises a $60M Series A for its agent authorization layer — OAuth-backed tool execution and credential management so agents act on users' behalf without god-mode access. 'What agents can do' joins execution and governance as a funded category of the harness stack.
OpenAI releases Symphony — open-source orchestration that turns project work into isolated, autonomous coding-agent implementation runs and surfaces the ones that pass, so teams manage outcomes instead of supervising sessions. Big Tech keeps raising the floor on what counts as a baseline harness.
Durable-execution leader Temporal closes a $300M Series D led by a16z at a $5B valuation, reporting 380% YoY revenue growth and 20M+ monthly installs. The agent-infra thesis in one round: in-memory agent state breaks in production, and durable execution is the production architecture.
ClickHouse acquires the open-source LLM observability platform Langfuse as part of its $400M Series D at a $15B valuation — the first big consolidation in agent observability. The open-source core remains maintained; the deal signals that agent tracing is becoming a database-scale problem.
Anthropic began enforcing its terms against third-party clients using Claude Code subscriptions, forcing alternative harnesses like opencode to rework authentication (its issue thread became the debate's ground zero). A defining platform-risk moment for the open-harness ecosystem: the model provider owns the login, the harness does not.
Anthropic announces plans to donate MCP to a neutral foundation, with new registry, auth and sandbox specs — cementing MCP as vendor-neutral infrastructure for the entire agent ecosystem.
Gemini 3 debuts with state-of-the-art agentic coding benchmarks, alongside Antigravity — a browser-based agentic development environment where agents plan, edit and execute tasks across editor and browser surfaces.
The AI-native editor's latest round makes it one of the fastest-growing software companies ever — a bellwether for how central agentic coding has become to the harness economy.
Anthropic makes Claude Code generally available and holds its first developer conference — with Claude Code for web, enterprise features, and hooks/MCP improvements central to agent workflows.
Bret Taylor's Sierra raises at a $10B valuation on the strength of its enterprise agent platform — validation that vertical agents can carry massive enterprise valuations.
Microsoft announces the public preview of its Agent Framework, converging AutoGen's multi-agent research lineage with Semantic Kernel's enterprise abstractions — a major consolidation in orchestration frameworks.
Cognition raises a round valuing the Devin maker above $10B — among the highest valuations for a pure agentic-harness company.
Anthropic open-sources the SDK behind Claude Code — the agent loop, tool use, subagents, memory and hooks — letting developers build their own Claude-powered agents on production rails.
OpenAI launches GPT-5, touting major gains for coding agents and reduced 'thinking' friction in workflows like Cursor, Copilot and ChatGPT — a model release explicitly aimed at the harness layer.
After Google licensed key Windsurf talent, Cognition (maker of Devin) acquires the Windsurf product and brand — consolidating two of the biggest names in agentic coding.
The 2025 World's Fair cemented 'agent engineer' as a job title — harness reliability, context engineering and evals dominated hallway track and stage alike.
The A2A protocol — backed by 50+ partners at launch — is donated to the Linux Foundation for neutral, open governance, standardizing how agents discover and talk to each other.
The open-source library for letting agents drive browsers becomes one of the fastest-growing repos in AI, spawning hosted products and a wave of 'web operator' agents.
Google launches ADK, a modular open-source framework for building agents with Gemini integration, multi-agent hierarchies and A2A support — Google's major entry into the framework wars.
OpenAI releases an open-source Agents SDK (successor to the Swarm experiments) with handoffs, guardrails and tracing, plus a Responses API purpose-built for agentic workflows.
Anthropic open-sources MCP, a standard for connecting AI assistants to data and tools. Within a year it was adopted by OpenAI, Google, Microsoft and most of the agent ecosystem — becoming the de-facto tool-connect layer of agent harnesses.