GitHub user sshiv012 created a discussion: Texera Agent: issues to prioritize 
for the release 1.4

The AI Agent panel has a few user-facing problems that make it hard to rely on. 
Below are the five we'd like to address in the next release (about 10 days 
out). Please share thoughts, pick-ups, or anything we've missed.

| # | Issue | Priority |
|---|---|---|
| 1 | Agent chats can't be reopened, and model selection doesn't respond | Must 
have |
| 2 | Agent panel shows agents that don't belong to you | Must have |
| 3 | Agents can get stuck in long loops and can't be stopped | Must have |
| 4 | Agents can only use a small set of operators | Should have |
| 5 | Adding new models requires a restart | Optional (stretch) |

---

### 1. Agent chats can't be reopened, and model selection doesn't respond
**What users see:** Existing agent tabs show a lock icon and won't open. On the 
"+ Agent" tab, clicking a model card does nothing, so "Create Agent" stays 
greyed out.

**Expected:** Users can always reopen their own chats. If a chat belongs to a 
different workflow, show a clear "Open that workflow" option instead of a dead 
lock. Clicking a model selects it, and the agent can then be created. Add a 
neat UX feature that guides the user, some users found it confusing that they 
have to first click the model for the create agent button to be enabled, it is 
not super intuitive.

### 2. Agent panel shows agents that don't belong to you
**What users see:** The panel lists agents created by other people (shown as 
locked tabs), and the agent count includes them. Agents also vanish when the 
service restarts.

**Expected:** Each user sees only their own agents, and conversations aren't 
lost on a restart. The persistence part might be difficult to achieve so we can 
plan this for future release. 

### 3. Agents are slow and expensive, can loop, and can't always be stopped
**What users see:** Sometimes an agent keeps adding or editing operators round 
after round without finishing. Pressing Stop doesn't always halt it and it can 
keep running after the tab is closed or the agent is deleted.

Even normal requests cost far more than they should. At every step the agent 
resends the whole conversation and workflow, rebuilt from scratch. Because of 
that, the model provider almost never reuses what it has already processed (a 
low prompt cache hit rate), so each step is billed as if it were the first. On 
top of that, the agent re-checks and re-runs the workflow after every edit. 
Cost and wait time grow quickly with each step, and a looping agent multiplies 
all of it.

**Expected:**
- Each request has a sensible (default) limit on steps, time and cost. When the 
agent hits a limit, it tells the user instead of silently continuing.
- Stop always works, and closing or deleting an agent ends its work.
- The agent reuses unchanged context between steps so cached input is used, and 
it skips repeated workflow checks and runs that aren't needed.
- Usage (tokens and cache hit rate) is visible per conversation, so we can see 
the improvement.

### 4. Agents can only use a small set of operators
**What users see:** The agent only builds workflows from a limited set of basic 
operators (file scan, filter, join, aggregate, a few charts, Python UDF). It 
can't use the rest of Texera's operator catalog, even when a better operator 
exists.

**Expected:** The agent can use all operators available in Texera. Users or 
admins can still narrow the list if they want to.

### 5. Adding new models requires a redeploy *(optional for this release)*
**What users see:** The model list in "Select Model Type" is fixed. Making a 
new model available requires an admin to change the deployment and restart 
litellm service. Picking a model that isn't actually usable only fails later, 
mid-conversation.

**Expected:** Admins can add or remove models from within Texera without a 
restart, and agent creation fails right away with a clear message if the chosen 
model isn't available.

> Marked optional: this release can ship without it if time is tight. But 
> please update the default config to reflect latest models at the time of 
> release. 

---

**Ask:** Please comment if you've seen other agent problems we should add.

GitHub link: https://github.com/apache/texera/discussions/9568

----
This is an automatically sent email for [email protected].
To unsubscribe, please send an email to: [email protected]

Reply via email to