AI agents move to the web, but verification becomes the real challenge
Anthropic, OpenAI, and Google focus on autonomous browsing, while Google Research releases Harness² to improve agents without modifying models. In both cases, the same limit emerges: web access does not guarantee data reliability.
Anthropic, OpenAI, and Google have recently launched agents capable of autonomously retrieving information from the web, each in their own way. But the real question is no longer just reaching a page: it is still necessary to know whether the content is up-to-date, complete, and trustworthy. This is where enterprise adoption is at stake, especially when the agent must handle prices, stocks, contracts, or compliance data.
Aller sur le web n’est pas comprendre le web
The three players converge on the same idea: agents must have their own path to the Internet rather than using a human’s browser. Anthropic integrated a browser directly into Claude. OpenAI abandoned its autonomous Atlas browser to consolidate browsing within ChatGPT’s work mode. Google, meanwhile, is advancing a cloud-hosted navigation agent. The timing differs, but the conclusion is identical.
Accessing a URL does not solve the underlying problem. A product page may display an outdated price, a site may serve different content to bots, and some information only appears after client-side loading. As a result, an agent may produce a very confident response... that is completely false. The bug is not noisy. That is precisely what makes it troublesome.
Le vrai goulot d’étranglement se déplace vers la confiance
The first reflex is often to oppose real-time browsing to indexed search. The first approach provides live access but is more expensive, slows queries, and faces bot detection. The second relies on a pre-built index, thus faster but also more quickly outdated. In both cases, data can be misinterpreted or already obsolete when the agent uses it.
This shift in the problem matters for the market. For two years, the priority was to make models reason through multiple steps. That part is progressing. Now, the limit is elsewhere: it is necessary to verify what the agent retrieves before letting it act. For companies, this is no longer a technical curiosity but an operational responsibility issue.
Google pousse Harness² pour améliorer les agents sans retraining
In another domain, Google Research has released Harness², an open-source framework under Apache 2.0 license that optimizes an agent’s environment without modifying the model’s weights. In other words, it acts on the skeleton around the model: system prompts, guardrails, memory, tools, sub-agents, scripts, and plugins. The model itself remains intact.
The framework relies on a two-step mechanism. First, the agent performs a task with its base harness. Then, an editor model proposes two variants of this harness: one corrects general errors, the other targets the relevant domain or procedure. Both versions are tested, then a composer keeps only what improves the observed result. Google presents this approach as a way to progress without perfect labels or exhaustive test sets. Again, the machine does not think better on its own; it is better supported.
Des gains réels, mais pas gratuits
According to results published by Google Research, a refinement step improves the four model-harness combinations tested on three professional benchmarks: WorkBuddy-Bench, JobBench, and Harvey’s LAB. The best gain reaches 11.8 percentage points on LAB, while the GLM-5 case under OpenCode rises from 64.2 to 75.7 on WorkBuddy-Bench, a +11.5 point increase. The smallest observed gain is 1.2 points. These figures remain first-party results, thus at this stage [to be verified] by independent replication.
The price to pay is not negligible. Each refinement adds executions, thus time and API costs. Google’s README also notes security warnings: unrestricted network access by default, sandbox disabled for Codex, and the need to define one’s own guardrails. The framework is designed for research, not for unsupervised production deployment. Agents can be very clever; they also like to self-reward if no one is watching.
What this means for companies
The future will depend less on web access speed than on the ability to prove that information is fresh. Teams deploying agents for support, purchasing, monitoring, or compliance will need to add a verification layer: cross-checking sources, detecting outdated data, logging changes. Without this step, the agent will remain fast but not necessarily reliable.
Google Research, like other labs, thus shows that the market is entering a more down-to-earth phase: one where an agent is no longer only asked to find an answer but to justify it. It is less glamorous than an autonomous browser. It is also much more useful.
Key points
- Anthropic, OpenAI, and Google launched web agents within days.
- Harness² targets prompts, tools, and guardrails, not model weights.
- Google announces up to 11.8 points gain on professional benchmarks.
- Real-time browsing remains exposed to outdated pages and anti-bot filters.
- Search versions require human controls and auditing.
- The key issue in 2025 becomes verifying web data.
By the numbers
- 11.8 points — maximum gain announced by Google Research on Harvey’s LAB.
- 11.5 points — GLM-5 improvement under OpenCode on WorkBuddy-Bench.
- 64.2 to 75.7 — GLM-5 score on WorkBuddy-Bench after one step.
- 3 benchmarks — WorkBuddy-Bench, JobBench, and Harvey’s LAB.
- 2 approaches — real-time browsing and indexed retrieval.