Daily AI Implementation Scout Council

Today's ranked top 20

#1

newbuild nowBoth runtimes31 / 35

Pydantic AI v2.42.0, cross-turn state and thinking-block fixes repo

Pydantic AI shipped v2.42.0 on September 8, 2026 and a further release on September 12, adding new model support and fixing a bug where Anthropic thinking blocks were not recovered correctly across turns in long agent sessions. This continues the framework's weekly release cadence from v2.38 and v2.39 earlier in the month.

What it does for you: YY already runs the typed-agents-pydantic-ai skill on this framework, so this is a same-day version bump, not a new adoption decision. The thinking-block recovery fix matters directly for any long-running Pydantic AI agent that uses Claude's extended thinking, since the bug could silently drop reasoning state across turns.

In practice: Routine but load bearing, the kind of fix you want applied quietly before it bites a real run.

For: Both runtimes. Plain Python framework callable from a Claude Code skill wrapper or directly inside Hermes' Python runtime, not tied to either host.

Security4

Quality5

Auditability4

Useful to you5

Useful to community4

Buildable now5

Hermes4

Verdict: build now. Already adopted infrastructure with a concrete correctness fix and low integration cost, this is an update to something in use, not a new build decision.

Build #1 Pydantic AI v2.42.0, cross-turn state and thinking-block fixes: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/pydantic/pydantic-ai/releases. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · v2.42.0 released September 8, 2026, with a further release September 12, 2026, on top of v2.38 and v2.39 earlier in the month.

#2

newtest firstBoth runtimes28 / 35

Claude Style Patch, a CLAUDE.md prose constraint file skill

Merged into awesome-claude-code on September 14, 2026, this is a CLAUDE.md section that bans specific prose habits: announcing a point before making it, colon-hinged sentences where the left side labels the right, verbless fragment openers, and stacked compression. Each rule names the habit, shows an example, and gives the rewrite.

What it does for you: YY already runs his own no em dash and plain-language prose rules in his global CLAUDE.md. This patch names several more habits worth checking his own writing against, and being a single file, it costs nothing but a read to compare against what he already enforces.

In practice: A small, sharply scoped resource, the kind that either adds two real rules or confirms YY's own rules already cover it.

For: Both runtimes. A markdown CLAUDE.md section works for any agent host that reads CLAUDE.md, not limited to Claude Code, though its current audience is Claude Code users.

Security5

Quality3

Auditability5

Useful to you3

Useful to community3

Buildable now5

Hermes4

Verdict: test first. Trivial to test, drop it next to YY's existing prose rules on one report and see whether it catches anything the current no em dash rule misses, before deciding whether to fold any of it in.

Build #2 Claude Style Patch, a CLAUDE.md prose constraint file: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/andrewroxby/claude-style-patch. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Merged as PR #2832 into awesome-claude-code on September 14, 2026, a single-file, single-contributor submission.

#3

newtest firstBoth runtimes26 / 35

E2B core SDK v2.49.1 repo

E2B's core SDK reached v2.49.1 on September 10, 2026, the latest point release on the most used open sandbox SDK for agent code execution.

What it does for you: If Axion or Hermes ever needs to run untrusted or generated code outside the main process, this is the SDK most likely to be reached for. A point release like this is worth a changelog skim before the next time sandboxed execution comes up, not an urgent action today.

In practice: Steady maintenance on infrastructure that matters when it is needed and is invisible until then.

For: Both runtimes. Python and JS SDK callable from either a Claude Code tool call or a Hermes worker process.

Security3

Quality4

Auditability3

Useful to you4

Useful to community4

Buildable now4

Hermes4

Verdict: test first. No current Axion workflow depends on E2B yet, but buildable_now and community adoption are both high enough that a quick trial is worth more than a watch-only note.

Build #3 E2B core SDK v2.49.1: use the ai-implementation-build-intake skill to build this safely. Source: https://pypi.org/project/e2b/. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · v2.49.1 uploaded to PyPI September 10, 2026.

#4

newtest firstBoth runtimes24 / 35

Stagehand v3.7.3, stable CDP-native release repo

Stagehand's v3 line moved from the branch release tracked August 28 to a stable published npm tag, v3.7.3, this week, with fixes to computer-use coordinate normalization and structured output provider handling. v3 dropped the Playwright dependency for a direct Chrome DevTools Protocol connection.

What it does for you: For any MCL or Axion task that needs an agent to actually click through a real website, such as funnel click throughs or form tests, Stagehand's CDP-native v3 is now stable enough to trial rather than watch from the sidelines.

In practice: A framework that has now shipped its architecture change and moved on to fixing rough edges, a healthier sign than the branch-only release tracked two weeks ago.

For: Both runtimes. Node and TypeScript SDK, runnable from a Hermes node process or invoked by a Claude Code browser task.

Security3

Quality4

Auditability3

Useful to you4

Useful to community4

Buildable now4

Hermes2

Verdict: test first. Directly relevant to render-verify style click-through work YY already requires, worth a real trial run against one funnel page before deciding whether to depend on it.

Build #4 Stagehand v3.7.3, stable CDP-native release: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/browserbase/stagehand/releases/tag/%40browserbasehq%2Fstagehand%403.7.3. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Tag @browserbasehq/stagehand@3.7.3 published this week, following the v3 branch release tracked August 28, 2026.

#5

newwatchBoth runtimes22 / 35

Vercel Sandbox v4.3.0, non-persistent sandboxes repo

Merged September 8, 2026, Vercel Sandbox v4.3.0 adds a non-persistent sandbox mode (sandbox sh --rm), region hints and failover defaults across all 19 Vercel regions, and Vercel Managed Images support.

What it does for you: A disposable, short-lived sandbox is cheaper and lower latency than a persistent one for one-off agent tool calls, closing a gap with E2B and Daytona. Worth knowing about if Axion ever needs quick, throwaway code execution close to wherever a funnel page is served.

In practice: A catch-up feature more than a leap, matching what competitors already offered.

For: Both runtimes. JS/TS SDK usable from a Node based Hermes worker or a Claude Code tool call, not tied to either host specifically.

Security3

Quality4

Auditability3

Useful to you3

Useful to community3

Buildable now3

Hermes3

Verdict: watch. No current Axion dependency on Vercel's stack for sandboxed execution, and the feature itself is parity with existing options rather than a differentiator, so this stays a watch item.

Build #5 Vercel Sandbox v4.3.0, non-persistent sandboxes: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/vercel/sandbox/pull/326. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · PR #326 merged September 8, 2026, versioned as v4.3.0.

#6

newwatchBoth runtimes22 / 35

Mem0 OSS hardening patches, week of September 8 repo

Mem0 landed three reliability and security fixes the week of September 8 to 15, 2026: it removed a silent pip auto-install in the litellm proxy path (now raises ImportError instead), fixed get_all() dropping the page query parameter, and moved a feature-notice check from a live PostHog flag call to a static fail-safe file.

What it does for you: These fixes matter only if you already run Mem0 in a restricted or sandboxed deploy, since the silent auto-install was a real supply-chain surprise. YY's own memory-tool-decision review already concluded Axion does not need a memory tool, so this is informational rather than actionable.

In practice: Sensible hygiene fixes on a tool YY has already decided to pass on.

For: Both runtimes. Open source Python memory layer, callable from either a Claude Code skill or a Hermes process.

Security4

Quality3

Auditability3

Useful to you2

Useful to community4

Buildable now3

Hermes3

Verdict: watch. Good fixes on their own terms, but useful_yy stays low because the prior memory-tool-decision review already closed this question for now.

Build #6 Mem0 OSS hardening patches, week of September 8: use the ai-implementation-build-intake skill to build this safely. Source: https://docs.mem0.ai/changelog. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Three dated fixes in the Mem0 changelog landing the week of September 8 to 15, 2026.

#7

newwatchBoth runtimes22 / 35

Milvus v3.0.1 repo

Released September 9, 2026, Milvus v3.0.1 adds REST v2 snapshot management, expanded reranking support, TEXT field support in the Go client and RESTful API, plus Storage V3 and data-consistency fixes.

What it does for you: If a future Axion or MCL retrieval build needs a self-hosted vector database instead of a managed one, the reranking and snapshot management additions directly affect retrieval quality and backup safety. Nothing in the current stack uses Milvus today.

In practice: A serious, incremental release on infrastructure that only matters once a real RAG build exists.

For: Both runtimes. Vector database server with client libraries in Python, Go and JS, not tied to Claude or Hermes specifically.

Security3

Quality4

Auditability3

Useful to you2

Useful to community4

Buildable now3

Hermes3

Verdict: watch. No current Axion or MCL project runs on Milvus, so this stays a watch item until a retrieval build actually needs a self-hosted vector database.

Build #7 Milvus v3.0.1: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/milvus-io/milvus/releases/tag/v3.0.1. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · v3.0.1 tagged and released September 9, 2026.

#8

newwatchBoth runtimes21 / 35

AG-UI Protocol release train, September 8 to 14 repo

AG-UI, the Agent-User Interaction Protocol, shipped five dated releases between September 8 and 14, 2026: fixes to resumed-run frontend tool handoff and thread-scoped working memory seeding, making CORS opt-in by default on createStrandsApp, tighter auth header forwarding with credential stripping, and token usage now populated in RUN_FINISHED and RUN_ERROR events.

What it does for you: This is the protocol layer for showing a running agent's progress live in a web page rather than only a finished result, relevant if a future here.now dashboard needs to stream an agent's work instead of polling for a finished state. The CORS default change is a breaking change for anyone already running an AG-UI integration publicly.

In practice: Active, near-daily shipping on a still pre-1.0 protocol, promising but not yet something to depend on.

For: Both runtimes. Protocol and SDKs are host-agnostic, usable from any frontend talking to any agent backend.

Security3

Quality4

Auditability3

Useful to you2

Useful to community4

Buildable now2

Hermes3

Verdict: watch. No current YY dashboard uses AG-UI, and the breaking CORS change argues for watching a few more release cycles before adopting.

Build #8 AG-UI Protocol release train, September 8 to 14: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/ag-ui-protocol/ag-ui/releases/tag/release%2F2026-09-14. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Five dated GitHub releases tagged September 8, 9, 10, 11 and 14, 2026.

#9

newwatchBoth runtimes20 / 35

modelcontextprotocol/servers memory server, concurrency fixes plugin

The official reference memory MCP server landed two correctness fixes on September 3, 2026: serializing graph mutations to prevent a concurrent-write race, and skipping duplicate entities and relations within a single batch.

What it does for you: This is the reference memory server that ships in modelcontextprotocol/servers, installable directly rather than built. The concurrent-write fix matters if two agent sessions ever write to the same memory graph at once, a real correctness bug, not a style change.

In practice: A small, precise bug fix on infrastructure meant to be trusted, exactly the kind of thing worth noticing.

For: Both runtimes. An MCP server is host-agnostic by design, installable into Claude Code, Hermes, or any MCP client.

Security4

Quality3

Auditability4

Useful to you2

Useful to community3

Buildable now2

Hermes2

Verdict: watch. No current Axion or Hermes workflow runs the reference memory MCP server, so this is tracked rather than acted on, though it would be a quick install if that changed.

Install: add npx -y @modelcontextprotocol/server-memory as an MCP server entry in .mcp.json or via claude mcp add. Review before enabling. Copy

Source · Two fix commits dated September 3, 2026: one serializing graph mutations, one deduplicating entities within a batch.

#10

newwatchStandalone tool19 / 35

Braintrust Patterns, autonomous failure-mode triage repo

Braintrust shipped Patterns this month, a feature that runs its Loop agent on a schedule to auto-detect recurring failure modes and cost regressions across traces, and proposes fixes.

What it does for you: Turns observability from manual dashboard reading into an autonomous triage layer, relevant if YY ever runs a production agent at a scale where nobody has time to read every trace by hand.

In practice: A genuinely new capability tier for evals tooling, not just a UI polish release.

For: Standalone tool. Hosted SaaS eval platform, YY would subscribe to it rather than embed its code.

Security3

Quality4

Auditability2

Useful to you3

Useful to community3

Buildable now2

Hermes2

Verdict: watch. No Axion project currently runs at a scale that needs autonomous trace triage, so this is a watch item pending that need.

Build #10 Braintrust Patterns, autonomous failure-mode triage: use the ai-implementation-build-intake skill to build this safely. Source: https://www.braintrust.dev/docs/changelog. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Feature dated to Braintrust's September 2026 changelog entry.

#11

newwatchStandalone tool19 / 35

Sourcegraph Amp, free BYOK tier and steer-while-running repo

Amp became free to use when bringing your own compute and model keys on September 13, 2026, following two usability releases on September 8 and 10: messages now steer the running agent immediately instead of queuing, and users can put custom agents behind Amp's model-selection Dial.

What it does for you: Removes the cost barrier to trialing a frontier-focused coding agent harness as a comparison point against Claude Code, useful context when evaluating tooling even without switching.

In practice: A pricing change plus real usability fixes, worth a look, not worth switching for.

For: Standalone tool. Closed-source coding agent CLI, used alongside Claude Code rather than embedded in it.

Security3

Quality3

Auditability2

Useful to you3

Useful to community3

Buildable now3

Hermes2

Verdict: watch. YY already standardizes on Claude Code, so this is a comparison data point, not an adoption candidate today.

Build #11 Sourcegraph Amp, free BYOK tier and steer-while-running: use the ai-implementation-build-intake skill to build this safely. Source: https://ampcode.com/chronicle. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Free BYOK tier announced September 13, 2026, following usability releases September 8 and 10, 2026.

#12

newwatchStandalone tool18 / 35

Arize AX, session-level Annotations repo

Arize AX shipped session-level annotations on September 8, 2026, letting teams send an entire multi-turn session or trace, not just a single span, to an annotation queue via the UI or the Python, TypeScript, or Go SDK.

What it does for you: Most real agent failures span multiple turns rather than one call, so this closes a real evaluation gap for anyone running multi-turn agents in production. No current Axion project is instrumented with Arize.

In practice: A sensible, overdue feature for how agents actually fail.

For: Standalone tool. Hosted SaaS observability platform, consumed as a service rather than embedded.

Security3

Quality4

Auditability2

Useful to you2

Useful to community3

Buildable now2

Hermes2

Verdict: watch. No current Axion telemetry runs through Arize, so this is informational for now.

Build #12 Arize AX, session-level Annotations: use the ai-implementation-build-intake skill to build this safely. Source: https://arize.com/docs/ax/release-notes/history/2026/09-2026. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Shipped September 8, 2026 per Arize's dated release notes page.

#13

newwatchStandalone tool17 / 35

LangSmith Cloud changelog, September 7 to 14 repo

LangSmith Cloud's weekly changelog for September 7 to 14, 2026 covers feedback in annotation queues now being attributed to whichever identity the request resolves to, and engine issue validation now retrying when preparing replay traces hits a transient error.

What it does for you: Minor correctness fixes to a platform YY does not currently run traces through. Worth noting as part of the ongoing LangSmith versus Langfuse comparison, not actionable today.

In practice: Routine weekly maintenance, nothing that changes the buy decision either way.

For: Standalone tool. Hosted SaaS observability platform tied to LangChain's cloud, consumed as a service.

Security3

Quality3

Auditability2

Useful to you2

Useful to community3

Buildable now2

Hermes2

Verdict: watch. No current YY project runs through LangSmith, so this is a routine tracking note.

Build #13 LangSmith Cloud changelog, September 7 to 14: use the ai-implementation-build-intake skill to build this safely. Source: https://docs.langchain.com/langsmith/changelog. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Dated changelog section for September 7 to 14, 2026.

#14

watchStandalone tool17 / 35

Cursor Projects (beta), persistent multi-agent workspace repo

Cursor launched Projects in beta on September 10, 2026, a workspace mode that maintains context over months of work and delegates tasks to many subagents under a coordinator agent that plans and checks work rather than writing code itself.

What it does for you: A step toward persistent, self-directed multi-agent project management rather than single-session coding help. Interesting as a direction, but it is a beta feature inside a closed-source editor YY does not use as the primary driver.

In practice: Bold framing, autonomy claims that need a real trial before believing them.

For: Standalone tool. Closed-source IDE feature, used as a separate coding environment rather than embedded in Axion or Hermes.

Security3

Quality3

Auditability1

Useful to you3

Useful to community4

Buildable now2

Hermes1

Verdict: watch. Beta status and closed-source auditability keep this at watch until independent reports confirm the coordinator-agent claims hold up.

Build #14 Cursor Projects (beta), persistent multi-agent workspace: use the ai-implementation-build-intake skill to build this safely. Source: https://cursor.com/changelog. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Launched in beta September 10, 2026, per Cursor's own changelog.

#15

newskipBoth runtimes17 / 35

docling-java v0.6.6 repo

docling-java, a Java port of the Docling document-parsing library, released v0.6.6 on September 14, 2026.

What it does for you: Would matter for a JVM-based document ingestion pipeline. Axion and Hermes are Python and Node, with no JVM stack anywhere, so this has no near-term use.

In practice: A solid release in a language YY's stack does not run.

For: Both runtimes. Java library, technically host-agnostic, though no JVM host exists anywhere in Axion or Hermes today.

Security3

Quality3

Auditability3

Useful to you1

Useful to community3

Buildable now2

Hermes2

Verdict: skip. The score alone would land in test-first range, but there is no JVM stack anywhere in Axion or Hermes to run this in, so fit disqualifies it rather than the score.

Build #15 docling-java v0.6.6: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/docling-project/docling-java/pull/683. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · v0.6.6 released September 14, 2026, tracked via PR #683.

#16

newwatchBoth runtimes17 / 35

Cline v4.1.17, shared SDK harness rebuild repo

Cline's harness was rebuilt around a shared Cline SDK pulled out of the VS Code extension, version-bumped September 2, 2026 with follow-on tags continuing through the second week of September.

What it does for you: An architecture change affecting stability and extensibility of a widely used free, bring-your-own-key coding agent extension. Worth a light comparison note, YY runs Claude Code as the primary driver already.

In practice: A major migration for an 11 million user base, real engineering weight behind it.

For: Both runtimes. Open source VS Code extension usable with any model provider, not tied to Claude specifically.

Security3

Quality3

Auditability2

Useful to you2

Useful to community4

Buildable now2

Hermes1

Verdict: watch. Significant migration for its own user base, but not a tool YY currently depends on, so this stays a tracking note.

Build #16 Cline v4.1.17, shared SDK harness rebuild: use the ai-implementation-build-intake skill to build this safely. Source: https://cline.bot/blog/how-we-migrated-11-million-users-to-clines-biggest-harness-upgrade. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Version bump dated September 2, 2026, described as migrating 11 million users, with follow-on tags through the second week of September.

#17

newwatchStandalone tool16 / 35

OpenAI Codex CLI v0.155.0-alpha cadence repo

Following v0.154.0 (already tracked September 10), Codex CLI has been shipping near-daily alpha builds toward v0.155.0, tagged alpha.3.9 on September 11 and alpha.4 on September 14, 2026.

What it does for you: Mostly a signal of pace, not a specific feature YY needs today. Worth noting only because OpenAI is iterating on its CLI coding agent multiple times a week right now, a useful data point when comparing coding-agent options later.

In practice: Rapid alpha churn, more useful as a pace signal than as something to adopt mid-cycle.

For: Standalone tool. OpenAI's own terminal coding agent, run side by side with Claude Code rather than embedded inside Axion or Hermes.

Security2

Quality3

Auditability2

Useful to you2

Useful to community4

Buildable now2

Hermes1

Verdict: watch. Alpha builds are not something to depend on yet, and Claude Code already covers this role for YY, so this is a watch-only competitive signal.

Build #17 OpenAI Codex CLI v0.155.0-alpha cadence: use the ai-implementation-build-intake skill to build this safely. Source: https://github.com/openai/codex/releases. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Alpha tags dated September 11 and September 14, 2026, following the September 10 v0.154.0 release.

#18

watchStandalone tool14 / 35

AWS Agent Registry, bedrock-agentcore namespace deadline repo

AWS Agent Registry moved to the agent-registry namespace on August 6, 2026, and AWS confirmed the old public-preview bedrock-agentcore namespace will be discontinued on September 17, 2026, two days after this issue.

What it does for you: Not an Axion or MCL dependency today, but the imminent deadline is worth flagging to anyone on YY's network running an early AgentCore integration, since a namespace cutover on a fixed date can break a pipeline with no warning if missed.

In practice: A hard deadline, not a feature announcement, the risk is entirely about the calendar.

For: Standalone tool. AWS managed service, consumed as a hosted registry, not embedded code.

Security3

Quality3

Auditability2

Useful to you1

Useful to community3

Buildable now1

Hermes1

Verdict: watch. No current YY project runs on AWS Bedrock AgentCore, so this is flagged for awareness rather than action.

Build #18 AWS Agent Registry, bedrock-agentcore namespace deadline: use the ai-implementation-build-intake skill to build this safely. Source: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/release-notes.html. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Namespace migration began August 6, 2026, with the old bedrock-agentcore preview namespace discontinued September 17, 2026.

#19

newwatchStandalone tool14 / 35

GitHub Copilot weekly releases, September 8 to 11 window repo

Three dated GitHub Copilot changelog entries this window: enterprise-managed sandbox for Copilot in JetBrains (September 8), enterprise-managed permissions letting admins centrally block, require approval for, or allow agent operations (September 9), and VS Code Agents added to Copilot usage metrics reporting (September 11).

What it does for you: None of these apply to YY's own Claude Code based workflow, but the enterprise-managed-permissions feature is a useful reference pattern: a centralized allow or block list for agent actions is the same shape as Axion's own build-gate and ghl-write-gate hooks, just implemented for a different platform.

In practice: Steady enterprise-control feature work, the kind that ships quietly every week.

For: Standalone tool. Hosted SaaS coding assistant tied to GitHub's own platform, consumed as a service.

Security3

Quality3

Auditability1

Useful to you2

Useful to community3

Buildable now1

Hermes1

Verdict: watch. YY does not run GitHub Copilot as a primary agent, so these are competitive-pattern notes rather than adoption candidates.

Build #19 GitHub Copilot weekly releases, September 8 to 11 window: use the ai-implementation-build-intake skill to build this safely. Source: https://github.blog/changelog/2026-09-09-enterprise-managed-permissions-for-github-copilot-agent-operations. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Three dated changelog entries: September 8, September 9, and September 11, 2026.

#20

newskipStandalone tool10 / 35

Zed agent panel, ACP reconnect without restart repo

Zed's stable channel release notes reference the Agent Panel gaining the ability to reload a broken external agent connection over ACP, such as Claude Code or Codex, without restarting the editor, alongside reduced memory use in long sessions, in the period following Zed's September 1 multiplayer-agent announcement.

What it does for you: A minor reliability fix for anyone plugging an external CLI agent into Zed's editor. YY does not use Zed as a working environment.

In practice: The weakest sourced item this issue, a real feature but without a single dated changelog entry pinned to it.

For: Standalone tool. Closed-source editor feature, used as a separate environment rather than embedded in Axion or Hermes.

Security2

Quality2

Auditability1

Useful to you1

Useful to community2

Buildable now1

Hermes1

Verdict: skip. The evidence here is a stable-channel release page rather than a specific dated changelog entry, the weakest sourcing of the batch, so this is skipped rather than padded into a watch item on thin ground.

Build #20 Zed agent panel, ACP reconnect without restart: use the ai-implementation-build-intake skill to build this safely. Source: https://zed.dev/releases/stable/latest. Save canonical skill/agent under AXION\Skills and AXION\Agents. Copy

Source · Referenced on Zed's stable release notes page, following the September 1, 2026 multiplayer-agent announcement, without a single pinned dated entry.