RunningTab: KAIST Overcomes Coding Agent Context Drift via Environment-Side Tabs for Direct Workspace Interaction

Background and the Problem

Autonomous coding agents increasingly operate via Direct Workspace Interaction (DWI)—using terminal commands to inspect repositories, read logs, and modify files without precomputed vector indexes. However, during long-horizon engineering tasks, agents suffer catastrophic context drift: bulky file outputs flush initial task constraints out of the context window, listed but uninspected files are forgotten, and agents prematurely declare victory while critical requirements remain completely untouched.

Architecture and How It Works

To eradicate agent amnesia, researchers from KAIST introduce RunningTab (arXiv:2610.10444), externalizing memory tracking into an environment-managed state machine:

1. Environment-Side Persistent Tab: Rather than relying on unreliable internal model scratchpads, RunningTab maintains an environment-hosted ledger tracking three structural pillars: Requirements (explicit task objectives), Provenances (file excerpts tagged with exact filepaths and line ranges), and Candidates (unopened file paths discovered via directory exploration).
2. Structured Tab Resolution: Prior to each reasoning turn, the environment surfaces the tab, enabling the agent to match gathered evidence against pending requirements or provide explicit justifications for deferred items.
3. Mandatory Finish Check Barrier: When an agent attempts to finish execution, the environment intercepts the termination token. If unverified requirements remain unresolved, the system blocks exit and prompts the agent to rectify outstanding criteria, eliminating premature terminations.

Benchmarks and Measured Results

Evaluated on demanding enterprise and software benchmarks including Workspace-Bench, TheAgentCompany, and OfficeQA Pro:

1. Outperforms Baselines Across All Benchmarks: RunningTab consistently beats standard DWI and self-refinement baselines across multiple frontier LLM backbones.
2. TCR@70 Jump: In TCR@70 (percentage of tasks fulfilling >= 70% of rubric requirements), RunningTab achieves 25.67, significantly exceeding the 20.67 achieved by DCI baselines and 22.33 achieved without environment-side tabs.
3. Internal Memory Failure Mode: Ablation studies confirm that storing tracking ledgers inside model context prompts causes performance to collapse back to vanilla baseline levels, proving that environment-level isolation and strict completion barriers are mathematically essential.

Getting Started for Developers

The RunningTab findings offer actionable architectural blueprints for coding agent developers. Rather than relying solely on expanding context windows, production systems (such as autonomous PR fixers and CI/CD agents) must deploy external environment-side tracking daemons. Intercepting shell operations to maintain provenance trees and implementing execution completion barriers fundamentally improves agent delivery reliability in real-world enterprise repositories.