agentmode.aiAsk for help
Open sectionsClose sections

The July 20 field edition

Agents are learning where to stop.

This week’s most useful releases made agents easier to inspect, contain, and trust. Here are the five changes worth knowing.

By Turwin Labs656 sourced records
Read the field notes Explore the field map
A physical miniature agent system links memory, tools, approval, artifact handling and handoff around a blue task route.
One task moves through memory, tools, approval, artifacts, and handoff. The blue route marks every place the work can wait, change, or pass to someone else. AI-assisted editorial illustration.

In this editionA live index of what you can read today.

656
sourced records
301
systems and frameworks
157
terms and concepts
2
records held for more proof

The week in the field

What made agents easier to trust this week?

Five releases changed what agents may touch, keep, and pass along. Each note begins with the official release record. If one touches your work, read the source. Better boundaries make an agent easier to use without giving it the run of the place.

01

Framework release

A blue route crosses three square checkpoints before meeting a black mechanical stop in a miniature architectural model.
A good trace shows the route, every checkpoint, and the place a person stopped the work. AI-assisted editorial illustration.

Pydantic AI adds capability hooks and trace controls

Pydantic AI 2.13.0 adds hooks for model-resolution capabilities. It also lets operators leave model-request parameters out of span attributes. That can reduce what traces record, but it does not make a full telemetry review optional.

Read the primary source
02

Security-relevant release

A cutaway architectural model shows one blue path moving through guarded black-and-bone chambers.
A file boundary has to hold from the first check to the place where the session writes. AI-assisted editorial illustration.

Strands Agents hardens file-backed sessions

Strands Agents TypeScript 1.10.0 fixes symlink attacks in FileSessionManager. It also adds intervention handlers and unified storage. The fix closes a specific hole in one component; other storage backends still need their own review.

Read the source
03

Security-relevant release

Blue tokens move through transparent and black test frames while one route diverts into a safe holding tray.
The valid path continues. The rejected action waits where someone can inspect it. AI-assisted editorial illustration.

OpenAI limits schema expansion and trace-error exposure

Agents SDK Python 0.18.3 limits $ref expansion during strict-schema conversion and hides non-tool details in trace errors. Both fixes narrow known risks. Neither proves that every agent or trace is secure.

Read the source
04

Framework release

Black-and-bone tool modules circle a selector that feeds one chosen module onto a blue execution rail.
A skill can be picked and reused like a tool. Someone still has to own its permissions and release. AI-assisted editorial illustration.

CrewAI takes its Skills Repository out of experimental status

CrewAI 1.15.4 moves its Skills Repository beyond the experimental label. That makes CrewAI’s own surface more settled. It does not guarantee that the repository will work with every other skills format.

Read the source
05

Security-relevant release

A blue command route enters a guarded black-and-bone artifact chamber through layered inspection gates.
The tool call crosses a confirmation gate before it can touch the artifact. AI-assisted editorial illustration.

Google ADK tightens tool confirmation and file access

ADK Python 2.5.0 blocks forged continuations in tool confirmation, rejects user-written function calls in resumable mode, and checks artifact paths for cross-user access. These are useful fixes to named problems, not a promise that every ADK system is secure.

Read the source
Read all field notes

The reference desks

Where do you want to begin?

Choose the desk that answers the question in front of you. The glossary explains a term. Systems compares frameworks. Ecosystem shows who is building. Standards shows the rules they share. Start where the thing you need to know is easiest to name, then follow the links.

Most read · 01–05

Where should you begin?

New to the field? Read these five in order. They cover the agent itself, the work it follows, the rules that hold it in bounds, the protocol it uses for tools, and the test that tells you whether it worked.

  1. 01AI agent
  2. 02Agent workflow
  3. 03Agent governance
  4. 04Model Context Protocol
  5. 05Agent evaluation

The visual report

Which record should you read next?

Start with any term, system, organization, or standard. The map shows the records sitting closest to it, then lets you open each source for yourself. Use those connections to choose what to read next. A line suggests a path; it does not prove that the two things belong together.

Explore the full field map
Drawing the field plate

The map will appear here when its records are ready.

From the field

Found a problem worth fixing?

Turwin builds the system when the answer needs more than a definition.

Keep the guide

Want a copy you can keep?

Get the glossary with its dates and sources.

Send this form only if you want a reply about the guide or your request. We do not sell the list, and a person decides how we respond.

Learn · Explore · Build

Where should you go next?

  1. 01

    AgentMode.ai

    Learn the agent field.

  2. 02

    TurwinLabs.com

    Explore ideas worth testing.

  3. 03

    Turwin.ai

    Build the venture.