# ChangeLog
This changelog highlights the **regular updates** to AgenticFlow, focusing on two key areas:
* β¨ **Improvements** β Enhancements to existing features, workflows, and overall usability.
* π **Fixes** β Bug fixes and stability updates to ensure a smoother and more reliable experience.
The changelog is **updated every Office Hours** when we have a release, so you can easily track what's improved and what's been fixed week by week.
π [**View Previous Changelog (Office Hours #1-28)**](https://www.youtube.com/@AgenticFlow/videos)
***
## 12 May 2026 - Office Hours #47: Model Selector Overhaul & Smarter Chat
**Overview:** This release reshapes the **Model Selector** β the surface every agent and chat session passes through β into a three-panel layout that keeps the provider list, model list, and tuning controls visible at the same time. Each model now surfaces its **Max Input / Max Output token counts** and inline **Vision** / **Reasoning** capability badges, so capability discovery happens before you commit to a conversation rather than after a silent failure. On the chat side, image uploads are now gated to Vision-capable models, `.json` attachments are supported, and file attachments render as cards instead of raw text. Plus a cluster of bug fixes that close out long-tail issues with file attachments and the Telegram publishing integration.
### ποΈ Model Selector β New 3-Column Layout
* **Three panels, always visible.** Provider list on the left, model list in the center, model settings (Temperature, Max Output, Max Input) on the right β no more stacking or collapsing. Compare models across providers and tune parameters without losing your place in the list.
* **Token limits surfaced inline.** Each model shows its **Max Input** and **Max Output** counts directly in the selector, so context-window questions get answered before you start the conversation.
* **Vision & Reasoning capability badges.** Two new inline icons appear next to every model name β π **Vision** for image input support and π§ **Reasoning** for extended chain-of-thought. Capability discovery is instant, especially for newer or less-familiar models.
* **Provider icons in the left panel.** Faster visual scanning when picking among the dozens of providers wired into AgenticFlow.
* **Auto-scroll to the active model.** Opening the selector now jumps the center panel to whichever model is currently selected, instead of dropping you at the top of the list.
### π¬ Chat Improvements
* **Image uploads gated to Vision-capable models.** The image attachment button is disabled when the active model lacks the Vision badge β eliminating the frustrating case where the UI accepted an image but the model silently dropped or errored on it at inference time. Switch to any π Vision model to attach images.
* **`.json` file attachments supported.** Pass structured data β API responses, configuration payloads, dataset samples β directly into a conversation without copy-pasting.
* **File attachments render as cards.** Attached files now appear as compact cards showing the file name and type label, replacing the previous raw-text rendering that was easy to lose track of in longer threads.
### π Bug Fixes & Stability
* **Agent chat no longer hangs on file attachments.** Attaching a file to an agent chat was causing the agent to stop responding entirely β no error, no reply, just silence. Fixed; agent conversations now continue normally regardless of whether a file is attached.
* **Unsupported file types and oversized files filtered before send.** The root cause of the related silent-hang bug was that the UI accepted files that were rejected silently downstream. Now, unsupported formats and over-limit files are caught at upload time with clear feedback β no more waiting for a response that never comes. **Supported attachment types** in agent chat: plain text, PDF, images (on Vision-capable models), JSON, and common document formats, within the per-file size limit.
* **Telegram publishing restored.** The Telegram publishing integration was down due to a platform-side outage. The connection has been restored; Telegram remains a fully supported publishing channel for agent outputs and workflow triggers.
***
## 28 Apr 2026 - Office Hours #46: Frontier Model Wave & Time-Aware Workspace
**Overview:** April brought a back-to-back cluster of frontier model launches β **Claude Opus 4.7** (16 Apr), **GPT-5.5** (23 Apr), and **DeepSeek V4 Pro + V4 Flash** (24 Apr). All four are wired into AgenticFlow within days of release, so you can swap them into existing agents and workflows without changing a line of configuration. On the engine side, this release closes out the timezone story we promised in #45's Stabilization Sprint, ships the advanced MCP auth options builders have been asking for, and removes Composio MCP after their upstream routing endpoints went dark.
{% embed url="" %}
### π§ New Frontier Models
* **Claude Opus 4.7 β the new coding leader.** Anthropic's April release jumps to **87.6% on SWE-bench Verified** (up from 80.8% on Opus 4.6) and **64.3% on SWE-bench Pro**, putting it ahead of GPT-5.4 (57.7%) and Gemini 3.1 Pro (54.2%) on real-world software engineering. Reach for Opus 4.7 as the default model on long-horizon agent loops, multi-file refactors, code review, and any workforce slot where the agent needs to plan, call tools repeatedly, and recover from its own mistakes. Same $5/$25 per-million pricing as Opus 4.6.
* **GPT-5.5 β the operator and tool-call champion.** OpenAI's late-April release dominates agentic execution benchmarks: **82.7% on Terminal-Bench 2.0** (vs. Opus 4.7 at 69.4% and Gemini 3.1 Pro at 68.5%), **78.7% on OSWorld-Verified**, and **98.0% on Tau2-bench Telecom** for customer-service workflows. Available via the standard OpenAI provider at $5/$30 per million with a 1M-token context window. Best when an agent has to drive a real shell, click through web environments, or run long customer-facing dialog trees.
* **DeepSeek V4 Flash on the AgenticFlow provider β the new default.** The 284B/13B-active model is available directly under our AgenticFlow provider with no extra setup, and its quality is already at the peak we'd previously have split between two tiers: **79.0% on SWE-bench Verified**, **91.6% on LiveCodeBench**, and reasoning strong enough to retire the old `deepseek-reasoner` slot. Use V4 Flash as your default DeepSeek model β chat, coding, reasoning, bulk classification, content generation, lightweight tool-using agents β there's no longer a reason to split.
* **DeepSeek V4 Pro on the PixelML provider β frontier-tier for specialized work.** New 1.6T-parameter MoE (49B active), lands as the **#2 open-weights reasoning model** on the Artificial Analysis Intelligence Index. The efficiency story: at 1M-token context, V4 Pro uses **\~27% of the inference FLOPs and \~10% of the KV cache** compared to V3.2. Reach for V4 Pro only when you have a specialized frontier-tier workload β extreme long-context reasoning, research-grade analysis, or evals where the very top of the open-weights leaderboard matters. For day-to-day work, V4 Flash is the right call.
* **Practical heads-up on DeepSeek V4.** DeepSeek's own evals show V4 Pro / V4 Flash answer with high confidence even when they don't know (94β96% answer rate on AA-Omniscience). Pair them with retrieval (Knowledge tab) or grounding tools when factual accuracy is load-bearing.
* **β οΈ Legacy DeepSeek models retiring 31 May 2026 (PST).** The previous-generation API model IDs β `deepseek-chat`, `deepseek-reasoner`, and the **DeepSeek V3.2 / V3.2 Speciale / V3.2 Exp** entries β will be fully retired and inaccessible after **May 31, 2026 (PST)**. Migrate every one of them to **DeepSeek V4 Flash** β its quality already covers both the previous chat and reasoner slots, so there's no need to split your fleet. Reach for **V4 Pro** only when you have a specialized frontier-tier workload that genuinely needs the larger model. The DeepSeek base URL and your existing API key stay the same β only the model field changes.
### π Global Consistency & User Experience
* **Platform-wide timezone standardization.** We've completely overhauled how time is calculated and displayed across the platform. Every timestamp now dynamically syncs with your browser's local timezone instead of defaulting to UTC β drastically reducing confusion when reading chat history, scheduled triggers, or task logs.
* **Precise chat history tracking.** Messages and activity logs in Chat History are now mapped to your exact local time down to the minute. When an agent runs a scheduled task, the time reflected in the chat matches your real-time expectation β no mental math, no UTC offset arithmetic.
* **Streamlined Tasks-tab timestamps.** Automated task execution logs in the new Tasks tab now mirror the same local timezone used in the chat interface. Cross-referencing "when did this trigger actually fire" against a chat thread is finally a single glance instead of a calculation.
### π Integrations & Security
* **Advanced MCP authentication protocols.** Building custom integrations is now more flexible: you can configure API key authentication directly via custom HTTP headers (e.g. `x-api-key: your_key_here`) rather than being locked into a single bearer-token shape. This unblocks every internal API and self-hosted MCP server that uses provider-specific header conventions.
* **Composio MCP deprecated.** To protect system integrity and avoid silent workflow failures, we proactively removed the Composio MCP integration after the provider deprecated their core routing endpoints without notice. If you had Composio-routed tools in a workflow, swap them for the equivalent direct integration β most are already available natively in the Marketplace.
### π Bug Fixes & Stability
* **Knowledge retrieval setting now persists on reload.** Fixed a state mismatch in Agent settings where toggling "Automatic Knowledge Retrieval" to ON would visually reset to OFF on a browser tab reload, even though the value was correctly saved server-side. The UI now reflects the true database state on every reload.
* **Workflow `.json` round-trip integrity.** Squashed a serialization bug affecting workflow migration. Workflows downloaded as `.json` from AgenticFlow now upload back into any other workspace cleanly β no syntax errors, no validation failures, no silently dropped node fields.
***
## 14 Apr 2026 - Office Hours #45: Core Stability, Traceability & Bug Fixes
{% embed url="" %}
**Overview:** This release is part of our **Stabilization Sprint** β a focused push to patch memory leaks, resolve multi-agent execution bugs, and surface the trace-visibility tooling needed to debug complex workflows. New-feature work on the engine is paused so these fixes can land cleanly. Everything below is already shipped and live in your workspace.
### π οΈ Engine Stability & Critical Fixes
* **Memory leak fixed in agent task cleanup.** HTTP clients weren't being closed after agent tasks finished, causing worker memory to grow monotonically and trigger unnecessary autoscaling. Patched. Expect a smoother, faster, more predictable execution environment β especially under sustained load.
* **Multi-agent metadata `workspace_id` misassociation fixed.** A long-tail bug where high-load runs could be misattributed to the wrong workspace is resolved.
* **Parallel-research workforce no longer hallucinates missing reports.** Agents were occasionally replying "awaiting report from Researcher Bβ¦" and halting. They now fire their tools independently, and the synthesizer produces clean prose with proper `(Researcher A)` / `(Researcher B)` attribution.
* **Workforce agents reliably use their attached tools.** Tool-calling is now enforced in workforce slot prompts whenever plugins are attached, eliminating cases where a plugin-equipped agent skipped its first tool call and returned a placeholder answer.
* **Workflow runs are forgiving and can block until complete.** `af workflow run` accepts both flat (`{"url":"β¦"}`) and wrapped (`{"input":{β¦}}`) body shapes. New `--wait` flag polls until the run reaches a terminal state and returns the final output inline β no more wondering whether a `null` output means "running" or "failed".
* **Public workforce URLs resolve again.** `af workforce publish` now returns the correct `/v1/workforce/public/...` paths (the old response had a stale prefix that 404'd) and includes a ready-to-paste `curl` snippet and a Web UI canvas link in the response.
### π Traceability & Configuration Safety
*Three features announced as "coming soon" in* [*Office Hours #35*](#13-jan-2026-office-hours-35) *are now fully live on the Web UI. If you saw the roadmap and wondered where they went β they shipped, and this is the follow-up callout we owed you.*
* **Agent Debug Mode (Trace View).** Every agent's thread now shows its full execution trace β every tool call, every decision, every intermediate output. Open any thread in the Web UI to see it. No more black-box failures; you can pinpoint exactly which prompt or tool call went wrong.
* **Task Dashboard.** A workspace-wide tracking dashboard in the side navigation. Status, duration, output, and failure reasons for every background task at a glance. Stop guessing which agent is still running.
* **Agent & Workflow Versioning.** Save snapshots of an agent's or workflow's full configuration (prompts, tools, plugins, graph). Roll back instantly if an update misbehaves. Publish specific versions for production. Iterating without fear of breaking what works is now first-class.
 
### βοΈ Platform & Infrastructure Optimizations
* **50 concurrent active schedules per workspace.** Up from the previous ceiling β overlap-heavy automation no longer bottlenecks.\
* **Standardized workflow inputs.** The workflow input schema is now unified across Build, Run, and Shared views so forms render consistently wherever users encounter them.
* **Suggested Replies now work for guest viewers of shared agents.** Unauthenticated visitors on a shared Agent link can see and use Suggested Replies to start conversations β one fewer friction point in the first-touch flow.\
* **API-key auth on Marketplace + MAS Workforce endpoints.** Full API-key authentication is now integrated on the Workflow Marketplace and Multi-Agent System Workforce endpoints β safe programmatic management of autonomous teams and marketplace items at scale.
* **Auto-refreshed model lists on OpenAI-compatible providers.** When you connect a custom OpenAI-compatible provider, the model list now refreshes before it's displayed β you always see the provider's current lineup.\
* **New models: Gemma 4 26B A4B IT and Gemma 4 31B IT.** Both available via the PixelML and AgenticFlow providers with full vision and tool-calling support.\
* **Custom HTTP headers on self-hosted and third-party connections.** Connect AgenticFlow to OpenAI-compatible endpoints that require auth headers, routing hints, or other custom headers. Backend is live; the frontend rolls out later this week.\
### π₯οΈ CLI & Desktop-Agent Contract
* **AgenticFlow CLI v1.10 β the Composition Ladder.** The three deploy verbs (`workflow`, `agent`, `workforce`) are now surfaced as rungs 0β6 on a single complexity ladder. Every blueprint ships with a `kind`, a `complexity` rung, and a ready `deploy_command` so AI operators and human developers pick the right primitive on the first try. Filter the catalog with `af blueprints list --kind --complexity `.
* **12 pre-built, tested blueprints β one command to deploy each.** Four workflows (`llm-hello`, `llm-chain`, `summarize-url`, `api-summary`), three single-agent blueprints (`research-assistant`, `content-creator`, `api-helper`), five multi-agent workforces (`research-pair`, `content-duo`, `api-pipeline`, `fact-check-loop`, `parallel-research`). Plugins pre-attached where applicable β no assembly required.
* **Live marketplace catalog from the CLI.** `af marketplace list/get/try` reads the unified backend catalog, auto-detects each item's type (agent / workflow / MAS), and delegates to the right clone flow.
* **AI Toolkit v4.3 routing.** Plugin packs for **Claude Code, OpenAI Codex CLI, Cursor, and Gemini CLI** now route through a shared composition-ladder decision rule. The CLI is the contract every desktop AI client drives β including **Ishi**, AgenticFlow's first-party desktop agent (launched as Claw in [Office Hours #34](#06-jan-2026-office-hours-34), rebranded Ishi in [Office Hours #35](#13-jan-2026-office-hours-35), under active tiger-team iteration).
* **New** [**Ecosystem overview**](/welcome-to-agenticflow/ecosystem.md) **and** [**Hands-On CLI Runbook**](/welcome-to-agenticflow/cli-walkthrough.md)**.** Two new top-level docs explain how the platform, CLI, Ishi, and AI Toolkit fit together, plus a 20-minute copy-paste walkthrough that builds a persistent demo stack (workflows + agents + a workforce) you can show customers.
* **Smoother long-running agents by default.** `recursion_limit` on blueprint-deployed agents is now **100** (was 25). Multi-step tool-heavy agents no longer silently exhaust their ceiling on real-world research and content tasks.
* **Cleaner default `af --help`.** Deprecated commands (`af pack`, `af paperclip`, `af company`) are hidden by default; set `AF_SHOW_DEPRECATED=1` if you still need to see them.
 
***
## 31 Mar 2026 - Office Hours #44
{% embed url="" %}
Improvements
* Workflow - Output image now has a download button instead of a static link.
* New models GLM-5 & GLM-5 Turbo. You can use these models via Pixel ML Provider or z.ai connection.
* New models Qwen 3.5 9B and Qwen 3.5 Flash under the AgenticFlow provider.
Fixed
* Fixed issue where cover photo on profile was not displaying.
* iPad screens are too small to split into 3 tabs like on PC, so the system prompt on iPad is now displayed here (see image).\
* Fixed issue where error icon for missing key connection was not showing on workflows included with downloaded agent templates.
* Fixed issue where template search results are now more accurate.
***
## 17 Mar 2026 - Office Hours #43
{% embed url="" %}
Improvements
Ishi - Sub Agent
**46 skills across 12 categories.**
Move Skills to side bar
Fixed
***
## 10 Mar 2026 - Office Hours #42
{% embed url="" %}
Improvements
* Nano Banana 2
* Agent Knowledge UI: When creating Knowledge in an Agent, the creation section will now display in Table format or Text format depending on whether the user selects Create in Table or Create in Text.
Fixed
* Fixed an issue where agents appeared as Unpublished even though they had already been published.
* Fixed an issue where "No triggers added" was displayed despite triggers being configured.
* Fixed a bug with the "Include Google Search" option in the Web Retrieval node.
* Removed outdated nodes: `Image_to_video V1`, `Image_to_video V2`, and `Image_to_video V3`.
***
## 03 Mar 2026 - Office Hours #41
{% embed url="" %}
Improvements
* Gemini 3.1 Pro
* Add option to enable Sub-agents directly in the Agent Editor settings (alpha version, still in testing, anticipate to have bugs)
* Allow users to customize the chatbot logo by updating the Agent avatar.
Fixed
* Fixed an issue where users could not upload images when accessing a Workflow via a shareable link.
***
## 24 Feb 2026 - Office Hours #40
{% embed url="" %}
Fixed
* Resolved an issue where published agents were not showing as **Published** in the UI due to a recent regression; status now updates correctly after publish.
***
## 10 Feb 2026 - Office Hours #39
{% embed url="" %}
***
## 02 Feb 2026 - Office Hours #38
{% embed url="" %}
Fixed
* Agent - System Prompt field:
* Fixed: Markdown toolbar editing was not working in the System Prompt field across browsers (Firefox, Chrome).
* Fixed: Editor preview no longer removes line breaks or leading spaces; indentation is now preserved correctly.
* Note: The {{ }} syntax for referencing variables in Agents is temporarily unavailable. We are actively working on a fix.
***
## 27 Jan 2026 - Office Hours #37
{% embed url="" %}
Fixed
* Workflow Audio Transcription Failure\
Fixed an issue where users could not create a workflow to transcribe an MP3 file. The workflow failed even when reduced to a single step using only the Audio Transcription node.
* Ask AI Node β AppConnection Error\
Fixed an error where the Ask AI workflow node failed due to AppConnection creation issues caused by an invalid UUID and missing required fields.
* Member Page Refresh Permission Error\
Fixed an issue where refreshing the Member page incorrectly resulted in a βPermission deniedβ screen.
* ChatV2 Image Upload β NoSuchBucket Error\
Fixed a ChatV2 issue where uploading images failed due to broken Google Storage image links and a NoSuchBucket error.
***
## 20 Jan 2026 - Office Hours #36
{% embed url="" %}
**New Tasks Page** - We are introducing a new Tasks page to better manage your workflow.
Improvements
* We introduce new Tasks page
Fixed
* Fixed an issue where chatbot agents could not be created using one-click setup
* Fixed text extraction failures in the workflow node
* Restored missing input placeholders in agent configuration
* Fixed an issue where opening text was not displayed when Chat V2 was activated
* Fixed an issue where duplicating templates resulted in empty agent templates
* Fixed an issue preventing MAS templates from being duplicated
* Fixed an issue where Export Data to File produced an empty PDF in workflow nodes
* Fixed an issue with the built-in Web Search workflow node
***
## 13 Jan 2026 - Office Hours #35
{% embed url="" %}
**Debug Tray & Ishi Launch** - Unveiling the Debug Tray for complete agent execution visibility and introducing Ishi, the rebranded Executive OS (formerly Claw).
Improvements
* New Debug Tray for tracing agent execution steps
* Ishi (Executive OS) Launch - rebranded from Claw
* AgenticFlow Credits integration in Ishi
* MCP Server Integration in Ishi
* Task Dashboard roadmap announced
* Agent & Workflow Versioning coming soon
***
## 06 Jan 2026 - Office Hours #34
{% embed url="" %}
**Claw Launch & Free Pro Licenses** - Meet Claw, the $60K AI Employee. Free Pro licenses distributed to LTD customers via Partner Portal.
Improvements
* Claw Desktop AI Assistant Launch
* Free Claw Pro Licenses for LTD Customers (up to $4,950 value)
* Agent Skills - install, create, and customize AI capabilities
* AgenticFlow MCP Server Integration with 3,000+ tools
* Skill Creator for building custom skills
* Trace Debugging & Closed-Loop Development
***
## 30 Dec 2025 - Office Hours #33
{% embed url="" %}
**Introducing Claw - AI Architect** - Unveiling Claw, the AI architect for building AI infrastructures, and discussing physical robotics integration.
Improvements
* Claw Meta-Agent for Voice-to-Infrastructure
* Physical Robotics Integration ($50 Robot)
* Vision for 2026
***
## 23 Dec 2025 - Office Hours #32
{% embed url="" %}
**GPT-5-2 vs Claude Opus 4.5** - A showdown between GPT-5-2 and Claude Opus 4.5, plus new concept video series and UGC image generation.
Improvements
* GPT-5-2 vs Claude Opus 4.5 Benchmark
* Concept Video Series with Notebook LM
* UGC Image Feedback Template
***
## 16 Dec 2025 - Office Hours #31
{% embed url="" %}
**DeepSeek V3.2, Video AI & Image Gen** - Introducing DeepSeek V3.2, video understanding with Gemini, and credit-based image generation.
Improvements
* DeepSeek V3.2 Integration (20-30x cheaper than GPT-5)
* Video Understanding with Gemini
* Image Generation using Credits
* New Copilot Ask Mode
* Chat V2 Beta
***
## 09 Dec 2025 - Office Hours #30
{% embed url="" %}
**Marketplace 2.0 & Opus 4.5** - Launch of the revamped Marketplace, Publication Management, and Claude Opus 4.5 integration.
Improvements
* Marketplace 2.0 Launch
* Publication Management Dashboard
* Claude Opus 4.5 Integration
* Robomotion MCP Collaboration
* Zero to One Build Sessions Launch
***
## 02 Dec 2025 - Office Hours #29
{% embed url="" %}
**Black Friday Recap & Nano Banana Pro** - Black Friday recap, Nano Banana Pro whiteboards, and RoboMotion MCP integration.
Improvements
* Nano Banana Pro Whiteboards
* RoboMotion MCP Integration
* New Task Page UI Preview
* GLM 4.5 Air Model
* AWS Qualified Badge
---
# Agent Instructions: Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.
Perform an HTTP GET request on the current page URL with the `ask` query parameter:
```
GET https://docs.agenticflow.ai/changelog.md?ask=
```
The question should be specific, self-contained, and written in natural language.
The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.
Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.