Friday, September 11, 2026

An image-handling skill for agent workflows. It looks like a small building block for teams that need repeatable visual outputs.

If your agents produce creative assets, a clear handoff between text and image work keeps the process reviewable.

skills.sh · Creative ops · Content

A discipline for writing down technical and product choices as they are made, rather than retrying to reconstruct them from chat logs.

A decision log makes a client agent easier to operate, debug, and hand over when the team changes.

skills.sh · Engineering ops · Reliability

A focused guide for structuring Go projects when an agent is creating or changing services. It makes architectural conventions available at the point where code gets written.

Useful for keeping an agent-generated service legible to the next engineer instead of accumulating a custom layout one task at a time.

skills.sh · Engineering · Code quality

A guide for agents working in Svelte 5 codebases. It puts framework practices next to the work so common missteps are less likely.

A targeted skill can prevent a coding agent from retrieving generic advice and leaving a framework-specific mess behind.

skills.sh · Front end · Build quality

A practical reference for turning a campaign brief into video-ad requirements an agent can follow. It narrows the gap between a creative request and assets that can actually ship.

Useful when campaign work reaches production and the agent needs guardrails around delivery requirements, not only ideas for the creative.

skills.sh · Creative ops · Marketing

A gate that checks links for phishing and malware before an agent opens them. It fits unattended browsing flows.

Add it before a browser agent follows lead-research links or opens untrusted attachment routes.

Smithery · Security · Risk control

An editorial-research tool for finding sourced angles and ranking topics before the content team writes.

It should help turn the daily from a list of links into a few ideas worth actually writing or testing.

Smithery · Content ops · Editorial

A connector that turns existing report templates into PDFs such as invoices, contracts, and statements from an MCP-compatible assistant. It focuses on producing a finished document from a controlled template.

Useful when client ops repeatedly turn structured work into PDFs and you want the agent to preserve the approved layout rather than invent one.

Smithery · Document ops · Client ops

A coordination layer for AI-native teams with initiatives, milestones, tasks, delegated work, and persistent organizational context. It is aimed at keeping execution connected to an operating plan.

Worth testing where client agents cross sessions and owners need a readable view of decisions, work in flight, and approval points.

Smithery · Coordination · Operator ops

Tencent/teamai-cli

A TCL for making teams AI-native. It is a fresh open-source option for teams trying to coordinate agent work across an organization.

No license · 841 stars today · 4078 total

JustVugg/colibri

A pure-C engine for running mixture-of-experts models on existing hardware by streaming experts from disk. It is one more sign that local inference constraints are moving quickly.

Apache-2.0 · 98 stars today · 27625 total

OpenAI's new Agents API brings agent building into its developer docs. Read the surface area carefully: the operational question is how tools, state, and evals behave under real workloads.

HN

Cognition introduced SWE-2 as a new coding model. The claim is worth testing against your own bug-fix, feature, and regression workloads rather than benchmarks alone.

HN

Anthropic describes campaigns that abused models for attacks. For anything that browses or takes action, the operational lesson is to keep permissions, logs, and escalation paths explicit.

HN

Every makes the case for evaluations that a regular team can actually use, not just research groups. The timing is right: agent quality without a test set is largely a feeling.

Every

A user shows Claude turning to-dos and follow-ups into a daily worksheet for a reMarkable device. It is a tiny, concrete example of an agent fitting into a human workday.

R/CLAUDEAI

A practitioner redeciscovers that the setup around a flash model can change the result more than expected. The useful lesson is to compare harnesses, not just model names.

R/LOCALLLAMA

A post examines OUI-1, a model trained to make custom interface elements via a specialized DSL. It offers a different path for prototyping uis without starting with raw React code.

R/LOCALLLAMA

The same open-ended game brief was handed to two coding assistants. It is a reminder to examine the workflow and build quality, not just the first screenshot.

R/CLAUDECODE

A discussion argues that agent budgets are an authorization problem, not only a cost control. The design question is who can grant, cut, or review spending power.

R/LANGCHAIN

Copenhagen's Seed Capital closed a €130 million fund for Nordic seed to Series A investments in fintech, cybersecurity, resilience, and AI-led B2B.

€130M · Fund V

Munich-based Furo raised €3.44 million to expand its industrial battery-storage software and enter more European markets.

€3.44M · Venture round

CloudNC secured a $20 million Series B extension for manufacturing automation, taking total funding to $128 million.

$20M · Series B extension

Bluecore Energy raised a $50 million seed round for a nuclear energy startup, following a $10 million pre-seed and a stealth launch two months ago.

$50M · Seed

AgentGrad tries to update prompts in multi-agent systems using interventions and feedback. It matters if your team wants to turn run history into a controlled improvement loop.

89 upvotes · Sep 8

Miles describes a verifiable reinforcement-learning post-training stack, with components intended to be clean and customizable. It is a useful infrastructure reference if you're building evaluations or finetunes.ing.

50 upvotes · Sep 8

T1 trains a 122B-parameter mixture-of-experts model to work in a real shell for long-task runs. It is a direct fit for agent setups that need tols, error recovery, and persistence.

37 upvotes · Sep 10

This preprint compares reusable skill packages with subagents on long-running tasks. It is a strong operator topic because the choice affects how much context, tooling, and review work gets repeated.

Sep 11