Repo-Local Skills Turn Agent Work Into Maintained Workflows
OpenAI's Agents SDK maintenance write-up shows how repo-local skills, AGENTS.md, and CI make recurring agent work repeatable.
Read note↗
Notes from the engineering bench
Technology notes
Technology notes
for people who ship.
Field notes from the A2Techify engineering team on AI, developer tools, and the systems around them.
55Field notes published
31Technology threads
0%Paywalled
Small tests. Primary sources.
Useful judgment calls.
Follow a thread
01
The journal
Recently tested,
Recently tested,
carefully explained.
Practical research on AI systems, developer tools, security, and the infrastructure that makes them useful.
Copilot Dynamic Workflows Make Agent Runs More Repeatable
GitHub Copilot dynamic workflows put multi-agent orchestration into reusable code, with limits, checkpoints, and resumable runs.
Transformers Can Now Run GGUF Models Without Leaving Python
Hugging Face is bringing llama.cpp-style GGUF quantized checkpoints into Transformers, starting with fast Apple Silicon inference for Qwen3.5 models.
Dynamic Workers Make Agent Sandboxes Cheaper To Throw Away
Cloudflare Dynamic Workers move agent code execution into short-lived isolates, which is a practical shape for sandboxing generated JavaScript with narrow capabilities.
Copilot Local Sandboxing Makes Agent Permissions Concrete
GitHub Copilot app local sandboxing turns agent permissions into per-project policy, which is the right default shape for developer machines.
gpt-oss-safeguard Makes Safety Policies Runtime Inputs
OpenAI's gpt-oss-safeguard release turns written safety policies into runtime classifier inputs, which is useful for teams that need adaptable Trust and Safety workflows.
TRL Async GRPO Can Sync LoRA Adapters Instead of Full Weights
TRL v1.14 adds adapter-only vLLM sync for AsyncGRPOTrainer, making distributed RL training less dependent on shared GPUs and full-weight transfers.
OpenAI's Agents API Turns the Harness Into a Product
OpenAI's public beta Agents API exposes the managed Codex harness, including sessions, sandboxes, compaction, tool search, and subagents.
GitHub Coverage Rulesets Are Now Scriptable
GitHub added REST API support for code coverage ruleset thresholds, making coverage gates easier to manage across repositories without clicking through settings screens.
AI Agent Benchmarks Need a Consistency Number
IBM's ALTK-Evolve team showed why average pass rates can hide unstable agent behavior, and how Pass^k gives builders a better reliability check.
02
How we work
Less hype.
Less hype.
More signal.
We test new technology in the same way we would evaluate it for a real project: run it, inspect the trade-offs, and keep the parts that hold up.
- 01Run itStart with a small, reproducible test.
- 02Pressure-test itInspect the cost, security, and operational limits.
- 03Write it downShare the commands, code, and judgment calls.