Temporal 2026 年調查：80.8% 開發者每日或更頻繁使用 AI 智能體，49.1% 在正式環境運行

80% Use Agents Daily, 49% Ship Them to Prod

📖 Read the full version with charts and embedded sources on AgentConn →

Temporal's State of AI Agent Development 2026 survey landed this week with a number that should make every engineering leader pause: 80.8% of developers now use AI agents daily or more, up from 47.3% a year earlier. That adoption curve looks like a rocket. But the second number is the one that matters: only 49.1% run those agents in production. The 31-point gap between daily use and production deployment is not a timing problem. It is a harness problem.

The same day that survey dropped, Harrison Chase took the stage at LangChain Interrupt London and said what the GitHub trending page has been screaming for months: "The harness is the moat. Own your intelligence." Not the model. Not the prompt. The system around the model — evals, memory, skills, governance, and durable execution — is what separates the 49% who ship from the 51% who demo.

This is not a niche opinion anymore. It is a capital allocation signal confirmed across every platform we track.

The 80-49 Gap by the Numbers

• 80.8% of engineers use AI agents daily or more (Temporal, N=554)
• 49.1% run agents in production
• 21.8% say they use agents "continuously"
• 79.8% say compute costs meaningfully limit progress
• 96% of devs do not fully trust AI-generated code (Sonar)
• 48% always verify AI code before committing (Sonar)

Harrison Chase: Own Your Intelligence

At Interrupt London, Chase laid out what he calls the "company harness" — the five components that turn an LLM API call into a production agent system. The framing is deliberately strategic: if the model is a commodity (and with Gemini 4 Carbon, Opus 5.5, and GPT-6 Astra all within striking distance of each other, it increasingly is), then the harness is where you build defensible value.

The five components Chase named:

1. Evals — automated tests that verify agent behavior against your specific domain. Not benchmarks. Your benchmarks.
2. Memory — persistent context that survives across sessions. What the agent learned yesterday should inform what it does today.
3. Skills — reusable capabilities (tools, prompts, workflows) that encode your team's domain knowledge.
4. Governance — audit trails, spending limits, approval gates. The infrastructure of trust.
5. Fine-tuning loops — feeding production data back into the model to make it better at your specific tasks.

Chase's argument on the Sequoia podcast sharpened the point further: the companies currently doing best at harness engineering are coding companies. Not AI labs. Not foundation model providers. Companies whose core business is building software, and who treat the harness as a strategic asset rather than a wrapper.

This is the thesis behind LangChain's pivot from "build chains" to "build harnesses" — and whether you use LangGraph or not, the pattern is winning.

GitHub Trending: The Capital Allocation Signal

If Chase's keynote is the theory, GitHub's trending page on October 30 is the empirical evidence. Of the 15 trending repositories, 12 are agent infrastructure:

Notice the pattern. Memory (claude-mem, hindsight, mempalace). Skills (superpowers, skills, anthropics/skills). Quality gates (impeccable). Token efficiency (caveman). Execution environments (lightpanda, OpenBot, openbot). Every one of these maps to a component in Chase's harness framework. The market is not waiting for a better model — it is building the scaffolding to make current models actually work in production.

The combined star count across these 12 repos exceeds 1.29 million. Superpowers and skills alone are pulling over 300K each. This is not a trend — this is infrastructure consolidation.

The Verification Bottleneck: Why Agents Stall at the Production Gate

The 80-49 gap has a name: the verification bottleneck. Agents generate code, decisions, and actions faster than humans can verify them. And without verification, there is no production.

Anthropic's Claude Code team published the most concrete data on this problem. Changes requiring review at Anthropic went from approximately 2,500 to over 30,000 per day in four months. Their solution was not to slow down generation. It was to build verification into the harness: run checks while Claude builds, not after. The result: 2x more changes shipped with approximately 40% fewer reverts.

View the full Sonar press release →

Sonar's developer survey tells the same story from the other side. 96% of developers do not fully trust AI-generated code. Only 48% always verify before committing. And 38% say reviewing AI code requires more effort than reviewing human-written code. The output multiplied. The verification capacity did not.

The Verification Gap

Sonar's Tariq Shaukat at AI Engineer World's Fair: lines of code are up dramatically, but releases are up only 30%. Human reviewers agree with wrong AI output 4 out of 5 times. The bottleneck is not generation — it is knowing whether the generation is correct.

This is the strongest evidence for Chase's harness thesis. If your harness does not include automated verification — evals that test agent output against your domain's ground truth — you are in the 51% that uses agents for demos and prototypes but cannot ship them.

The Production Risk Nobody Priced In

The verification gap is about correctness. But production agents face a second, more dangerous gap: trust in autonomous action.

At the same AI Engineer conference, SoFi's Ishween Kaur demonstrated six attacks on x402 agent payments — the emerging standard for agent-to-agent financial transactions. In each attack, an agent paid for nothing. The question Kaur posed is one that no amount of model intelligence can answer: how does an agent decide whether to pay five cents?

David Sacks drew the line that production teams are now navigating:

View original post on X →

The answer is governance — the fourth component of Chase's harness. Spending limits. Approval gates. Audit trails. Per-action authorization. These are not model features. They are harness features. And without them, the 49.1% production number will not move.

Community Reaction: The Skeptics Have a Point

View the full thread on Hacker News →

The Hacker News thread "AI isn't replacing SREs, it's deskilling them" pulled 466 points and 341 comments. The top comment drew an analogy to aviation: "Just as the FAA mandates manual flying time even when the autopilot is perfectly capable, SRE teams can designate certain incidents as human-practice opportunities." The parallel to agent harnesses is direct — production agents need human-in-the-loop checkpoints not because the agent cannot handle the task, but because the organization needs to maintain the capability to catch when it fails.

View the full thread on Hacker News →

A separate thread on "Building an AI-native engineering org" drew sharper criticism. The most-upvoted comment: "Not a single mention of value or revenue. Only three metrics went up: LOC, token spend, and presumably release speed. What's the markup?" This is the vanity-metrics trap that Chase's harness framework is designed to escape. Evals tied to business outcomes — not lines of code, not tokens consumed — are what distinguish a production harness from a demo wrapper.

Contrarian Corner: The Harness Tax

Before you rush to build a five-component harness, consider the counter-argument. Portkey's analysis measured the "Harness Tax" — the fraction of each request that is harness overhead rather than useful work. Their finding: harnesses that were load-bearing in January became dead weight by March as models improved. An ICML 2026 paper on "The Scaffold Effect" found that harness choice can produce large performance differences — but that the effect is shrinking with each model generation.

Theo's provocatively titled "Are Harnesses Dead?" asks whether the category will exist in a year.

But even Theo's own data contradicts the headline. The honest answer: probably, but in a different form. The harness layer is not dead — it is moving from prompt engineering and chain-of-thought scaffolding into infrastructure concerns (durability, observability, governance) that do not go away when models get smarter.

The bet is that the production gap is a systems engineering problem, not a model intelligence problem. If you believe models will solve their own reliability, observability, and auditability within 6 months, skip the harness. If you believe those are infrastructure problems that persist regardless of model capability — and the evidence from Temporal, Anthropic, and Sonar suggests they are — then the harness is where your competitive advantage lives.

View the full Portkey analysis →

The Architecture That Is Winning

The production pattern that emerged from multiple AI Engineer talks and Temporal's own documentation is converging on a two-layer architecture:

Layer 1: Workflow Engine (Temporal, Inngest, Restate)

• Durable execution — if the process crashes, it resumes where it left off
• Retry policies with exponential backoff
• Long-running coordination (hours, days, weeks)
• Audit logs and replay

Layer 2: Agent Framework (LangGraph, OpenAI Agents SDK, CrewAI)

• LLM orchestration and tool calling
• Prompt management and chain composition
• Memory and context management
• Eval integration

The Temporal LangGraph plugin makes this layering explicit: LangGraph handles the agent logic, Temporal handles the production guarantees. Neither layer alone closes the 80-49 gap. Together, they address the three production requirements that daily-use agents ignore: recoverability (the agent crashed — now what?), observability (what did the agent do and why?), and governance (who authorized this action?).

What This Means for You

The 31-point gap between daily use and production deployment is not going to close because models get better. It will close when teams build the harness to make agents trustworthy.

Your Production Readiness Checklist

1. Audit your harness against Chase's five components. Evals, memory, skills, governance, fine-tuning loops. If you are missing two or more, you are likely in the 51% that cannot ship to production.

2. Measure verification, not generation. Your agent's value is not how much code it writes. It is how much verified, correct output ships. Track reverts, rollbacks, and post-deploy incidents per agent-generated change.

3. Evaluate the Temporal/LangGraph pattern. If you are building agents that need to run for more than a few minutes, survive crashes, or handle real money, you need durable execution. The framework-inside-workflow-engine pattern is becoming the default.

4. Budget for governance infrastructure. Spending limits, approval gates, audit trails. SoFi's x402 attack research shows that agent autonomy without governance is a liability, not a feature.

5. Watch the GitHub trending signal. When 12 of 15 trending repos are harness components, the market is telling you where the next wave of tooling maturity is happening. Adopt early or build late.

The model war is a spectator sport. The harness war is where builders compete. And right now, the scoreboard says 49.1%.

For more on the agent infrastructure shift, see our analysis of why memory became the new moat and which tools agents actually reach for. For the demo-to-production gap, read why 90% of agents die after the demo.

Originally published at AgentConn