AI news often arrives stripped of the detail developers actually need: what changed, what is usable now, and where the claim stops. Here is a tighter roundup of eight updates across agent protocols, coding tools, open-weight models, workplace research, and security.
The cutoff for this review was July 30, 2026 at 01:54 BJT. Most of the source announcements landed on July 27–28, so this is a dated snapshot—not an attempt to repackage two-day-old news as “just released.”
1. MCP’s July 28 specification makes the core protocol stateless
The 2026-07-28 MCP specification removes protocol-level sessions, Mcp-Session-Id, and the initialization handshake. It also adds mandatory server/discover, while Tasks becomes a formal extension.
That should make MCP easier to run on standard HTTP infrastructure. It is also a breaking migration across clients, SDKs, and servers. A published specification does not mean today’s implementations already support it.
2. .NET agents can discover Skills from MCP servers
Microsoft’s experimental Microsoft.Agents.AI.Mcp package lets .NET agents discover SKILL.md files and ZIP or TAR skill bundles through skill://index.json.
The practical appeal is centralized capability management: an organization can update a skill without redeploying every downstream agent. The trust boundary still matters. Agents should connect only to trusted MCP servers, and scripts inside an archive are not executed automatically.
3. Cursor launched a ₹649/month India-only plan
Cursor Start India supports INR billing through UPI and cards. The plan includes Composer, cloud agents, mobile remote control, MCP, Hooks, and Skills.
The interesting part is not another feature checklist; it is regional pricing and payments becoming part of the agent-product stack. The limit is equally important: Start is available only in India, is not the same as the full Pro plan, and does not promise unlimited use.
4. A tuned static harness beat adaptation more often than expected
A new preprint on agent harness optimization compares 120 tasks, 4,620 trajectories, and 729 harness configurations. Under the tested budgets, a static baseline optimized with DSPy matched or outperformed Contextual Bandit and REINFORCE approaches, often with fewer tokens.
The useful lesson is to tune the fixed harness before adding online adaptation. But this is an unreviewed preprint, not a universal production rule, and its results should not be generalized beyond the evaluated tasks and budgets.
5. Kimi K3’s full weights are available—and very large
Moonshot AI’s Hugging Face repository exposes 96 safetensors shards totaling roughly 1.56 TB. The official GitHub repository describes a model with 2.8 trillion total parameters, 104 billion active parameters, and a one-million-token context window.
“Open weights” does not mean easy local deployment on an ordinary computer. The custom license also should not be treated as equivalent to an OSI-approved open-source license.
6. OpenAI’s workplace study points to wider job boundaries
OpenAI analyzed more than 800,000 work-related messages and found that 16.8% involved tasks associated with another occupation. After excluding broadly shared tasks, the cross-occupation share rose to 43.5% among occupation-specific messages.
That is evidence of fewer handoffs and broader task ownership, not proof that entire jobs have already disappeared. The study describes changing work patterns inside the observed data; it does not establish economy-wide job replacement.
7. Microsoft detailed a multi-agent cyber defense system
Project Perception Agent coordinates red-, blue-, and green-team agents around a vulnerability-management loop. Microsoft reports that MAI-Cyber-1-Flash reached 96% on CyberGym—12 percentage points above Mythos—at roughly 50% lower cost.
Those are vendor-reported results in a specific benchmark setting. The public preview is scheduled for August 3, so this was an announcement, not general availability at the time of this roundup.
8. GitHub added stronger Agent governance to JetBrains
The latest GitHub Copilot for JetBrains update adds OpenTelemetry export for Agent workflows, token ceilings for bring-your-own-key models, and one control for enabling or disabling built-in models. Claude Agent workflows can also use MCP servers and custom agents.
This is mainly a cost, permission, and observability upgrade. It should not be mistaken for a sudden improvement in the underlying models.
Taken together, these updates show agent tooling becoming more operational: protocols are simplifying state, skills are becoming distributable, and IDEs are adding the controls teams need after a demo becomes a system. The caveats are part of the news, not fine print.
AI-assistance disclosure: This roundup was drafted with AI assistance from a source-checked brief. Its claims, dates, caveats, and primary links were retained for human review; the author remains responsible for the final text and publication.
Top comments (1)
The MCP shift to stateless protocol behavior, alongside mandatory server discovery and Tasks as an extension, shows how quickly agent infrastructure is moving from experimentation toward compatibility work. The JetBrains additions-OpenTelemetry export, token ceilings, and a single switch for built-in models-feel equally important because teams eventually need to explain cost and permissions, not just produce a clever demo. For founders and engineers, the durable advantage may be operational clarity: a smaller, tuned harness with visible limits can be more valuable than adding another adaptive layer or chasing the newest model benchmark.