Trending AI Briefing: Friday, August 21, 2026 (afternoon ET)
For two years the coding agent lived in one person's terminal. Today it moved into the room where the team already argues about the work. GitHub shipped Copilot into Slack and Microsoft Teams within an hour of each other this morning, both spawning a dedicated channel per agent session that anyone can watch and redirect. The top of the GitHub charts is the same shape: DeepSeek's agent harness, a community plugin wiring it into nine chat platforms, and CopilotKit's OpenBot giving every AI coworker its own container and its own channel. What none of them fixed is the meter. An open Codex issue that surfaced on Hacker News overnight documents roughly $1,386 of estimated four-day Bedrock spend from one developer working alone.
What's hottest in AI news right now
GitHub Copilot in Slack entered public preview on August 21, and the interesting part is not the chatbot. Mentioning @GitHub in a channel or thread starts a cloud agent session that can triage bugs, investigate failures, implement changes in a secure cloud sandbox, and open a pull request. GitHub is also a launch partner for Slack Code, a new channel type designed for agents: Copilot spins up a dedicated code channel where the team follows the plan, inspects diffs, previews HTML artifacts, and can stop the session outright. GitHub's stated payoff is pedagogy rather than throughput, which is a strange and interesting thing for a changelog to claim. Working in the open means developers can watch how their teammates prompt the agent. Issues and pull requests carry the Copilot app identity, and repository admins can require an extra approval before anything attributed to that identity merges. (GitHub Changelog)
GitHub Copilot in Microsoft Teams shipped the same day with nearly identical mechanics and one different permission rule: anyone in the conversation can add context and help plan, but only participants with write access to the repository can trigger a code change. The pitch is the standup, hand an action item to the agent while the meeting is still running. Billing is the detail builders should read closely. Cloud agent sessions consume AI credits under usage-based budgets, and cloud sandbox usage is billed separately with its own product-level or SKU-level budget. Two meters, two budgets, one @-mention. (GitHub Changelog)
deepseek-v4-flash-vision-exp landed in DeepSeek's API docs and hit the Hacker News front page on August 21 with just under 400 points. The limits tell you what it is built for: up to 600 images per request, 8192 pixels per side, and a hard ceiling of 384 tokens per image because everything gets resized to roughly 800 by 800 before inference. A 2000-pixel image and a 5000-pixel image cost the same. It answers on the OpenAI-compatible endpoint, the Responses API, and an Anthropic-compatible /messages endpoint, the standard three-way posture for a model that wants to drop into an existing agent loop without a rewrite. (DeepSeek API docs, HN thread)
DeepSeek Harness sits at number two on Trendshift's daily board with a roughly 4.2k single-day star gain, against 181k total stars and 20k forks. It is DeepSeek's own open-source agent harness, MIT licensed, shipped as npx @deepseek-ai/dsh web and built on Cordis with an everything-is-a-plugin architecture. The README says developer preview and warns of compatibility-breaking changes in capital letters, which is honest and also the reason to read the star count as attention rather than adoption. (GitHub, Trendshift)
A Codex issue about Bedrock cache-write spend reached Hacker News overnight under the headline "10x charges," and the actual report is narrower and more useful than that. Issue #37674, opened August 9 against Codex CLI 0.147.0 on the native amazon-bedrock provider, says native Bedrock requests to GPT-5.6 Sol cannot opt into explicit prompt caching, because the request types carry neither prompt_cache_options nor prompt_cache_breakpoint. The reporter's Cost Explorer estimate across August 5 through 8 covers 3,656 requests, 171.94M cache-write tokens, and roughly $1,386 total, with cache writes at about 85% of model spend. One local session logged 76 requests averaging 88K cache-write tokens each and zero cached input tokens. The reporter is explicit that these are usage-derived estimates and not invoice amounts, and does not claim every cache write is a defect. OpenAI has it assigned, filed as a feature request rather than a bug. (openai/codex#37674)
CopilotKit OpenBot turned up as a new 2026 Trendshift entry with about 2.1k stars, the open-source version of the idea GitHub just shipped commercially. Each AI coworker gets its own Docker container, its own Chromium with its own logins, its own /workspace volume, and a channel of its own. Every browser, file, and MCP call routes through one gateway that resolves the target from a server-held snapshot, evaluates CEL policy, writes an audit row, and only then acts. Deny beats allow, a missing policy permits nothing, and a broken rule refuses rather than opens. It speaks AG-UI, so LangGraph, Mastra, CrewAI, Pydantic AI, and Google ADK agents plug in unchanged. Labeled alpha. (GitHub)
New tools and features worth actually trying
Slack Code channels with @GitHub are the cheapest way to find out whether shared agent sessions help your team or just create a new place to talk past each other. Start with triage, not code changes. Honest tradeoff: it needs Copilot Business or Enterprise, it burns your existing entitlement, and a channel where anyone can start a session is a channel where anyone can start a session at three in the morning.
npx @deepseek-ai/dsh web gets a full agent harness with a web UI on localhost:3080 in one command, and the plugin architecture is worth studying if you are designing your own extension surface. Honest tradeoff: developer preview with promised breaking changes, so anything you build against it this month is a prototype, not a deploy path.
OpenBot's audit and boundaries pages are the part to steal even if you never run the rest. Point a bot at a form, then open /admin/audit and read what was permitted, refused, and failed, with the rule named on every refusal. That is the shape an agent audit log should have. Honest tradeoff: alpha software that wants Docker, Bun 1.3+, PostgreSQL, and a CopilotKit Intelligence license before it starts, and it ships with OPENBOT_DEV_NO_AUTH admitting every request as an administrator.
deepseek-v4-flash-vision-exp for screenshot-heavy loops is worth a benchmark this week because of the 384-token cap. If your agent reads dashboards or UI captures, a flat per-image ceiling makes cost modeling trivial in a way resolution-scaled pricing does not. Honest tradeoff: exp means what it looks like, the downscale to roughly 800 by 800 will lose small text in dense screenshots, and images only work in user messages.
Trending AI repos on GitHub today
Star totals are from GitHub badge data pulled this run; daily deltas are Trendshift momentum scores, which are rankings rather than verified counts.
openai/codex: terminal coding agent, roughly 111k stars. Number one on Trendshift today, the same day its Bedrock billing issue hit the HN front page.
deepseek-ai/deepseek-harness: DeepSeek's plugin-architecture agent harness, 181k stars. Number two, roughly 4.2k gain.
obra/superpowers: agentic skills framework and development methodology, around 275k stars. Still leading the AI-skills topic, itself third on Trendshift's daily topic board.
mattpocock/skills: skills straight out of one working engineer's .agents directory, around 228k stars. New in 2026 and climbing fast.
CopilotKit/openbot: governed AI coworkers, each with their own container and browser, about 2.1k stars. The open-source answer to today's Slack and Teams launches.
volcengine/OpenViking: self-evolving context database unifying agent memory, knowledge RAG, and skills, around 32k stars.
earendil-works/pi: unified LLM API, agent loop, TUI, and coding agent CLI in one toolkit, around 95k stars.
Tencent/AI-Infra-Guard: red-teaming across agent scan, skills scan, MCP scan, infra scan, and jailbreak evaluation, about 5.2k stars.
xmanrui/dsh-im: connects DeepSeek Harness to nine chat platforms including Slack, Telegram, Discord, and WhatsApp, around 400 stars. Small repo, exactly today's pattern.
What actually matters from today's signal
Track the channel, not the model. Every launch on this board assumes the unit of agent work is a shared, observable session with a named identity and a permission boundary, not a private process on a laptop. That reframes four things builders treat as separate problems: agent identity becomes an org-chart question, audit becomes a product surface instead of a log file, permissions get checked per participant rather than per user, and the review gate moves from "did a human read the diff" to "did a human who is not the requester approve the bot's PR." GitHub shipped that last one twice today, which is a strong hint about what enterprise buyers actually asked for.
The counter-signal is the meter, and nobody addressed it. Copilot in Teams draws on two separate budgets a channel member can start spending with one mention. Codex on Bedrock burned an estimated 172M cache-write tokens in four days for a single developer because the client had no way to declare a cache breakpoint. Per-seat pricing was a mental model for tools that idle when you stop typing. A shared agent channel does not idle. It runs while you are in a meeting, which is precisely the feature both changelogs advertise. The first serious incident here will not be a prompt injection. It will be a finance conversation about a Slack channel nobody thought to disable over a long weekend.
Name what nobody shipped: not one of today's chat integrations gives you a per-channel spend ceiling or a session budget visible before you start. You get org-level budgets and after-the-fact reports. Turn any of this on next week, set the org budget first, and check it Monday.
Source access notes: The Claude Code and Codex changelogs both exceeded the fetch size limit and their saved tool-result files were not reachable from the shell, so this run leans on GitHub's changelog, vendor blogs, DeepSeek's API docs, Trendshift, and the Hacker News Algolia API. Product Hunt and arXiv were skipped given the strength of the vendor-blog signal. Star totals came from shields.io badge JSON, since api.github.com is blocked at the sandbox egress allowlist. GitHub's rendered HTML pages served stale counts for the fastest-moving repos on this board, so every figure here is the shields.io value re-pulled at the end of the run.