#ai#culture#team

AgentScope 2.0.10: the enterprise agent breaks when reading tables

AgentScope 2.0.10 reviewed: Word tables, duplicate Excel headers, and TTS. Why agents fail on input data and how to shorten time to value.

  • What is included in the release
  • Why a table breaks RAG
  • Voice: the same contract problem
  • What This Means for Business

Trigger

  1. Team AgentScope, an open-source multi-agent framework from Alibaba engineers, released version 2.0.10. It contains almost no new features.

  2. Two fixes concern how the RAG module reads Word and Excel tables; three more concern speech synthesis and recognition.

  3. This release interests me more than major announcements because it shows where an agent breaks in a real company.

  4. In my experience, it most often breaks at the input stage while parsing a document.

  5. The quality of this analysis determines how many weeks it will take for the agent to start delivering value.

What is included in the release

The version 2.0.10 changelog contains five items worth reading carefully: - #2961: RAG preserves Word table columns that the parser previously skipped; - #2948: RAG correctly handles duplicate and empty column headers in Excel; - #2980: TTS audio is labeled with the format in which it was actually synthesized; - #2990: TTS follows speech instructions provided in a specific call; - #2925: audio transcription from DashScope (Alibaba's cloud model platform) arrives as text.

The documentation also announced SOP scenarios (the agent follows a predefined standard procedure step by step), TeamPipeline for agent teams, and support for MiniMax models. The first two matter more for business.

Why a table breaks RAG

RAG (retrieval-augmented generation) works as follows: the system splits documents into fragments, converts them into vectors, and finds the closest passages for a question, which the model uses to write an answer.

The model sees only what the parser passes to it.

If the parser corrupts a table, the model will confidently answer based on corrupted data.

Let's look at two typical failures

In a Word contract, the table contains merged cells, and the parser drops a column: values shift, and the price ends up under "delivery time." In a supplier's Excel price list, two columns are called "Price" (wholesale and retail), while a third has no header.

If rows are assembled into a dictionary by column name, the second "Price" overwrites the first, and the unnamed column disappears.

The system reports no error, and the agent gives the manager the wrong figure.

Fixes #2961 and #2948 address precisely this class of errors.

Since these issues had to be fixed in a patch release, framework users are likely feeding it real corporate files and encountering such failures in production.

Assess where AI can deliver impact in your process

clients@kt.team Telegram @kt_team_it

Voice: the same contract problem

The TTS fixes look minor, but they follow the same pattern

If audio is labeled as one format but recorded in another, the player in a call center or mobile app will not play it.

If synthesis ignores the instruction "speak slowly and formally"

, the bot will voice a legally important phrase in an upbeat advertising tone

In both cases, one component violated its agreement with another, so the user encounters a failure even though the model itself worked correctly.

What This Means for Business

In a demo, the agent answers using a clean PDF, and the result is impressive

In practice, it receives 1C exports with merged cells, price lists from twenty suppliers in twenty different formats, scanned contracts, and specifications edited by three departments for years.

Here is my project-based hypothesis; I have no measurements: the team spends most of the time before the first result preparing input data, while model selection takes a small share.

This leads to a simple rule for managers.

The agent project metric is TTU, time to use: the number of days until an employee receives a correct answer from the agent based on real documents.

Until the test set includes your own “messy” files, the demo proves nothing.

If no one measures answer accuracy, a manager spends the budget on theater, and a specific person is ultimately held responsible for it.

How we address this technically

  1. At KT.Team, we start with the input data and only then choose the model. Normalization before vectorization.

  2. We collect product attributes in a PIM system (Akeneo, Pimcore), where each field has a type, unit of measure, and place in the taxonomy.

  3. The agent reads an already-cleaned product record, so it never receives the raw price list with two "Price" columns.

  4. Facets and product record text are built on the same layer, so the same data supports both search and agent answers. Structured access instead of exports.

  5. When data lives in 1C, we give the agent MCP access (a protocol through which a model calls external tools) to specific methods: "stock by SKU" and "order status."

  6. The model then does not read the Excel export and cannot misread it. A regression suite of ugly files.

  7. We collect real customer documents with merged cells, empty headers, and duplicate columns.

  8. We test every parser or framework update, including patches such as 2.0.10, against this set before deployment. A gateway between the agent and the models. LLM & Security Gateway routes requests between GigaChat, YandexGPT, Qwen, and Claude, logs responses, and masks personal data.

  9. The model can be replaced without rewriting the agent, and any answer can be reconstructed during error analysis.

Verdict

AgentScope 2.0.10 is a boring release, and that is exactly why it is useful. The maintainers are fixing the issues that affect real-world deployments: table columns, Excel headers, and audio formats. For a company choosing an agent framework, I recommend looking for precisely these fixes in the changelog: they say more about maturity than a list of supported models. An agent that reads your price list correctly will start making money sooner than an agent with a stronger model that reads the same price list with a one-column shift.

Discuss the article: AgentScope 2.0.10: enterprise agent…

Enter your email or phone number so we can get back to you.

Send via:

By clicking the button, you accept the offer and grant consent to the processing of personal data.

Articles on the topic

A collection on development management

Related videos

All videos
4:48 The Culture of Digital Transformation and Why It MATTERS Digital transformationCulture
9:17 Corporate culture / 70% of IT projects die because of this?! Culture
5:03 99% of Documentation Is GARBAGE?! How to Create Documentation That Actually Works Culture

News on the topic

All news