Pranav’s Weekly Commit

Cloudflare built the agent cloud—and OpenAI opened the cyber layer

Hey —

This gap ended up being mostly about infrastructure: where agents run, which models they use, and how much access we give them once they can touch real systems.

A few releases were easy to ignore. These weren’t.

The big one: Cloudflare’s agent stack

Cloudflare’s latest Agents Week was less one launch and more an attempt to cover the entire lifecycle.

The easiest way I found to break it down:

• Build and run

• @cloudflare/computer picks an execution environment for each task.

• Cloudflare Agents adds tracing, replay, and human approval.

• Programmable CI/CD can run code-defined pipelines and stage agent-generated repairs for review.

• Control what agents can do

• Cloudflare Wallets adds programmable payment infrastructure.

• The Agent Access Model focuses on identity and authorization.

• WriteGuard adds fine-grained controls around MCP servers.

• Make the web agent-readable

• WebMCP exposes structured website capabilities.

• Kitesurf is an agent-first browser running inside V8 isolates.

• MCPv2, AI Search, and the combined Workers AI/Gateway control plane cover discovery, retrieval, and inference.

My read: Cloudflare wants to become the infrastructure underneath long-running agents—not another agent framework.

Several pieces remain previews. But the stack itself is becoming clear.

Open the Agents Week recap →

https://blog.cloudflare.com/agents-week-review-august-2026/

DeepSeek completed the V4 rollout

DeepSeek released V4-Flash in public beta on July 31, then shipped V4-Pro GA across its app, web, and API on August 13.

Both models now support:

• 1M-token context and up to 384K output

• Native OpenAI Responses and Anthropic API compatibility

• Low, high, and max reasoning effort

• Agent-focused post-training

Flash is the cheaper, faster path. Pro targets harder production-agent workloads. DeepSeek also announced peak/off-peak API pricing beginning August 16.

My read: DeepSeek is shaping V4 around long-running coding agents, not merely releasing another chat model. Native Responses API support—and explicit Codex adaptation—makes this unusually practical.

Official announcement →

https://api-docs.deepseek.com/updates/

Model and pricing details

https://api-docs.deepseek.com/quick_start/pricing/

OpenAI opened frontier cyber models to approved partners

OpenAI expanded Daybreak so approved partners can provide governed cybersecurity services using its frontier models.

There are two layers:

• Daybreak Blue — regular defensive work: vulnerability discovery, secure-code review, threat modeling, incident response, malware analysis, and patch validation.

• Daybreak Red — separately approved specialist access for controlled exploit reproduction, penetration testing, red teaming, and more advanced security research.

This is not blanket access. Approval remains tied to the exact identity, organization, project, model, and product surface.

My read: this is partly a model announcement, but the larger change is distribution. Frontier cyber capability can now reach customer environments through vetted service providers—with authorization and human review built into the access model.

Read the official Daybreak documentation →

https://learn.chatgpt.com/docs/cyber-safety

Gemini 3.7 Flash landed with unusually aggressive pricing

Google released gemini-3.7-flash to GA on August 13, positioning it as its most capable Flash model for coding, agents, and multimodal reasoning.

Useful details:

• 1M-token input context

• 64K-token output

• Text, images, video, audio, and PDF input

• Function calling, code execution, search grounding, structured outputs, and preview computer use

• Low, medium, and high thinking levels

Introductory API pricing runs through December 31:

• $0.75/M input tokens

• $3.75/M output tokens

• Batch/flex: $0.375 input and $1.875 output

Those prices double on January 1, 2027.

Release notes →

https://ai.google.dev/gemini-api/docs/changelog#08-13-2026

Model details

https://ai.google.dev/gemini-api/docs/models/gemini-3.7-flash

Pricing

https://ai.google.dev/gemini-api/docs/pricing#gemini-3.7-flash

Two more execution-layer releases

NVIDIA Nemotron 3.5 Lightning + NeMo Switchyard

A 30B open MoE model paired with an open router that selects models per workflow step. NVIDIA claims up to 4× faster output and 30% faster task completion; those remain vendor benchmarks until tested on your workload.

NVIDIA’s announcement →

https://blogs.nvidia.com/blog/nemotron-lightning-switchyard-rtx-dgx/

GPT-5.6 Sol Ultrafast

OpenAI’s Cerebras-powered API preview claims up to 750 output tokens/second, or roughly 14× normal speed.

See the preview →

https://openai.com/index/previewing-ultrafast

Effect v4 finally reached RC

Effect v4 moved from beta to release-candidate status on August 12.

The first RC was 4.0.0-rc.108; effect@rc now points to 4.0.0-rc.109. The normal effect@latest tag still installs v3, so migration needs an explicit RC tag:

bun add effect@rc

This is still pre-GA, but RC is a meaningful line: API shape should now be much closer to final, and the broader Effect ecosystem has moved onto coordinated v4 RC releases.

My read: good time for migration experiments and library compatibility work. Still too early for a casual production upgrade without typechecking and testing the whole application.

Effect v4 RC release →

https://github.com/Effect-TS/effect/releases/tag/effect%404.0.0-rc.109

GitHub’s answer to dependent agent work

GitHub also launched stacked pull requests in public preview.

Stacks give dependent changes layered diffs, cascading rebases, CLI support, and bottom-up merging—without forcing reviewers through one giant PR.

It does not solve every coordination problem, but it removes a lot of manual branch maintenance.

GitHub announcement →

https://github.blog/changelog/2026-07-30-stacked-pull-requests-are-now-in-public-preview/

My earlier write-up →

https://pranavb.xyz/blogs/merging-in-parallel

Roles worth opening

All seven pages were live when rechecked on August 14.

•

FurtherAI — Senior Backend/Full-stack Engineer

https://jobs.ashbyhq.com/furtherai/fdbf6b32-97e5-49a7-9296-d7f38f265b1a

Remote India

₹60–90L + equity.

•

Raven — Backend/AI Engineer

https://www.ycombinator.com/companies/raven/jobs/7OMJHg2-software-engineer-backend-ai

Bengaluru

industrial data, retrieval, orchestration, and agents.

•

Bespoke Labs — Product Engineer

https://jobs.ashbyhq.com/bespokelabs/3751a4cb-bb4d-4113-afe6-9c87341f6739

Bangalore

full-stack systems for AI training and execution.

•

Sarvam — Backend Engineer, Chanakya

https://jobs.ashbyhq.com/sarvam/86ae80f8-b7eb-43a4-afde-fef58e77e23e

Bengaluru

India-focused AI platform work.

•

Handshake — Software Engineer II, RL Environments

https://jobs.ashbyhq.com/handshake/da7bc8fc-4b7a-410a-aaa1-2faafed7c9fc

India

model-training and evaluation environments.

•

Lighthouz AI — Software Engineer

https://www.ycombinator.com/companies/lighthouz-ai/jobs/05KYNGI-software-engineer-4-6-years-experience

Delhi/Gurgaon

document-heavy AI agents for freight accounting.

•

Flagright — Senior Software Engineer

https://jobs.ashbyhq.com/flagright.com/59717db2-4b4c-4867-949e-6634c4022589

Bangalore

AI-powered financial-crime infrastructure.

Openings move quickly, but all seven pages were live when I checked.

That’s it for this commit.

View the commit history or subscribe →

https://newsletter.pranavb.xyz/

— Pranav

PRANAV'S WEEKLY COMMIT

#002 · AUG 2026

LEAD STORY · CLOUDFLARE AGENTS WEEK

Cloudflare built the agent cloud—and the model layer moved with it.

Agents, cyber access, models, Effect v4 + jobs. Weekly-ish.

Hey —

This gap ended up being mostly about infrastructure: where agents run, which models they use, and how much access we give them once they can touch real systems.

A few releases were easy to ignore. These weren’t.

01 · THE BIG ONE

Cloudflare's agent stack

Less one launch—more an attempt to cover the entire agent lifecycle.

Build + run — @cloudflare/computer, tracing, replay, human approval, and programmable CI/CD.

Control access — Wallets, Agent Access Model, and WriteGuard for MCP permissions.

Make the web agent-readable — WebMCP, Kitesurf, MCPv2, AI Search, Workers AI, and AI Gateway.

My read: Cloudflare wants to become the infrastructure underneath long-running agents—not another agent framework. Several pieces remain previews, but the stack itself is becoming clear.

Open the Agents Week recap →

02 · MODEL RELEASE

DeepSeek completed the V4 rollout

CONTEXT

384K

MAX OUTPUT

V4-Flash — public beta July 31; cheaper, faster path for high-volume agent work.

V4-Pro — GA August 13 across app, web, and API; aimed at harder production-agent workflows.

Both — native Responses + Anthropic APIs, low/high/max reasoning, and agent-focused post-training.

My read: DeepSeek is shaping V4 around long-running coding agents, not merely releasing another chat model. Explicit Codex adaptation makes this unusually practical.

Official announcement → · Pricing

03 · CYBER ACCESS

OpenAI opened frontier cyber models to approved partners

Daybreak lets vetted providers deliver governed cybersecurity services using frontier models.

DAYBREAK BLUE

Vulnerability discovery, secure-code review, threat modeling, incident response, malware analysis, and patch validation.

DAYBREAK RED

Separately approved access for controlled exploit reproduction, penetration testing, red teaming, and advanced security research.

This is not blanket access. Approval stays tied to identity, organization, project, model, and product surface.

My read: the larger change is distribution: frontier cyber capability can now reach customer environments through vetted service providers—with authorization and human review built in.

Read the Daybreak documentation →

04 · WORTH TESTING

Gemini 3.7 Flash gets aggressive on price

$0.75 IN · $3.75 OUT

Per million tokens through December 31. Prices double January 1.

Context — 1M input, 64K output.

Inputs — text, images, video, audio, and PDFs.

Agent tools — function calling, code execution, search grounding, structured output, and preview computer use.

My read: worth testing for high-volume coding agents and subagents where frontier pricing becomes hard to justify. Scheduled price increase belongs in any long-term comparison.

Release notes → · Model · Pricing

05 · RELEASE ROUNDUP

Four more worth opening

OPEN MODEL + ROUTER

NVIDIA Nemotron 3.5 Lightning + NeMo Switchyard

30B open MoE plus workflow router. NVIDIA claims up to 4× faster output and 30% faster task completion; still vendor benchmarks. Open →

RAW SPEED

GPT-5.6 Sol Ultrafast

Cerebras-powered API preview claiming up to 750 output tokens/second, roughly 14× normal speed. Open →

TYPESCRIPT · PRE-GA

Effect v4 finally reached RC

effect@rc now points to 4.0.0-rc.109; normal effect@latest still installs v3. Good migration-test point, not a casual production upgrade. Open →

SOURCE CONTROL

GitHub stacked pull requests

Layered diffs, cascading rebases, CLI support, and bottom-up merges for dependent work. Announcement → · My write-up

06 · FRESH OPPORTUNITIES

Seven opportunities worth opening

Direct official application pages. All seven were live when checked on August 14.

FurtherAI — Senior Backend/Full-stack Engineer

Remote India · ₹60–90L + equity

Production AI agents for insurance workflows. Apply →

Raven — Backend/AI Engineer

Bengaluru · ₹20–45L + 0.1–1% equity

Industrial data, retrieval, orchestration, and grounded agents. Apply →

Bespoke Labs — Product Engineer

Bangalore · Hybrid

Full-stack systems used for AI training and execution. Apply →

Sarvam — Backend Engineer, Chanakya

Bengaluru · On-site

Backend work on an Indian full-stack AI platform. Apply →

Handshake — Software Engineer II, RL Environments

Bengaluru · In-office

Realistic environments for frontier-model training and evaluation. Apply →

Lighthouz AI — Software Engineer

Delhi/Gurgaon · ₹20–35L · Hybrid

Document-heavy AI agents for freight accounting. Apply →

Flagright — Senior Software Engineer

Bangalore · On-site

Core infrastructure for an AI-powered financial-crime platform. Apply →

Openings move quickly—verify status before applying.

That's it for #002.

If a role closes or a link goes stale, reply and I'll correct it in the next commit.

View the commit history or subscribe →

— Pranav

Pranav's Weekly Commit

Cloudflare’s agent cloud, new models, developer releases, and fresh roles

Cloudflare built the agent cloud—and the model layer moved with it.

Cloudflare's agent stack

DeepSeek completed the V4 rollout

OpenAI opened frontier cyber models to approved partners

Gemini 3.7 Flash gets aggressive on price

Four more worth opening

Seven opportunities worth opening

Vercel’s agent stack, a 50%-cheaper model claim, and a few more things