Know Today — July 22, 2026

AI agents Agent skills Developer tools Workflow automation Evals MCP connectors Browser automation Git hooks Access control

👀 Keep an eye on

Versioned skill packages make reusable agent workflows practical by combining instructions, references, assets, and safe helper scripts.
The useful boundary for agents is shifting from chat assistance to scheduled or event-driven jobs with explicit checks and approval points.
Git hooks can make agent review available at natural delivery points, but advisory use is the sensible first deployment.
Read-only connectors and constrained internal APIs are a more workable starting point than broad write access across business systems.
Browser-agent loops can accelerate UI refinement when screenshot artifacts, test criteria, and final review are managed deliberately.

Pattern	Practical implication
Progressive context	Keep global guidance light, then load task-specific skills only when needed instead of stuffing every session with documentation.
Cross-tool workflows	Evaluate an agent on an end-to-end job—such as preparing for a meeting or triaging logs—rather than on isolated answers.
Self-correction loops	Tests and measurable acceptance criteria let an agent revise its output before returning it, reducing repeated manual review.

Good first candidates: release-note drafts, commit summaries, incident-log triage, and visual-regression checks.
Package stable internal commands behind narrow, documented interfaces rather than exposing an unrestricted production environment.
Separate research permissions from write-capable actions, and require human confirmation for consequential changes.
Keep shared operational knowledge outside any single prompt so approved tools can use the same maintained source of truth.

Opinion: The surrounding harness—context, tools, automation, and evals—matters more to recurring agent work than small differences in model choice; use this to prioritize workflow design and observability in tool evaluations.
Opinion: “Could a capable teammate succeed with the context and permissions this agent has?” is a useful diagnostic; it directs fixes toward missing documentation, access, or acceptance criteria.

Verify current behavior: connector availability, hooks, scheduling, and tool permissions vary by product, plan, and environment.
Passing an eval demonstrates only what the eval measures, not broad correctness, security, or fitness for a high-impact action.
Claims about overnight autonomous optimization are anecdotal unless the benchmark, controls, and reproduction details are independently checked.

Build one portable skill. Put a recurring task’s instructions, a short reference, a template, and a harmless helper script in a version-controlled folder; the expected outcome is a repeatable starting context instead of a fresh prompt each time.
Make a review loop measurable. Ask an LLM to draft release notes from a small set of commits, then check for required headings, ticket references, and prohibited terms before you review the result; the outcome is a concrete baseline for whether automation saves time.
Use separate worktrees for parallel changes. Create two Git worktrees, run local app instances on separate ports, and have separate LLM sessions plan and implement a tightly scoped change; the expected outcome is parallel exploration without agents sharing a working directory.
Prototype a UI check with cleanup. Use browser automation to capture a before-and-after screenshot for one local page, store only the approved comparison, and remove transient captures; the outcome is a controlled visual-feedback loop rather than a growing pile of artifacts.