---
title: "Limited-context model reliability"
description: "In DrowAI 0.4.0, models with limited effective context may not reliably follow the complete agent system instructions."
section: "Troubleshooting"
updated: 2026-08-17
tags: ["troubleshooting","models","context-window","v0.4.0"]
source: "/user-guide/known-issues/limited-context-model-reliability"
---

# Limited-context model reliability

In DrowAI 0.4.0, models with limited effective context may not reliably follow the complete agent system instructions.

> **Affected version: DrowAI 0.4.0.** Lower-capacity models and provider tiers
> with limited effective context are not currently reliable for agent runs.

This limitation has been observed with **GPT-OSS** configurations and
**Mistral free-tier** models. DrowAI supplies the model with system
instructions, task context, capability definitions, tool schemas, execution
history, and tool output. Together, these can place more pressure on the
available context than these model configurations can handle reliably.

When the effective context is insufficient, the model may:

- omit or stop following parts of the DrowAI system instructions;
- lose important task state during a longer run;
- produce incomplete or malformed tool calls; or
- return an answer without completing the expected agent workflow.

The model may still generate a response, so this can look like an application
or tool failure even when the task runtime is operating normally.

## Current guidance

Use a supported model with a larger effective context window and reliable
instruction following for DrowAI 0.4.0. If a task behaves unexpectedly, run the
same task with a higher-capacity supported model before reporting it as a DrowAI
runtime issue.

Reducing the visible task prompt may help in isolated cases, but it does not
remove the internal context required by the agent and is not a reliable
workaround.

## Status

Compatibility improvements for smaller-context models are planned for a later
release. Future work will focus on context budgeting, prompt compaction, and
clearer model compatibility checks. There is no release date for this work yet.