Coding agents get managed like teams
Switching between unfinished tasks costs coding agents 11.5 points of performance, a new approval system checks commands by their effects, and harness tuning now trades quality against cost explicitly.
- In a preprint, the Switch-SWE benchmark found interleaving unfinished tasks causes an average 11.5 percentage point drop in resolution performance across seven models. Self-reported m:…4598
- In a preprint, Caper is a consequence-aware approval mechanism that pre-executes agent shell commands in sandboxes and judges them by observed effects such as file changes and network attempts. Self-reported m:…5951
- In a preprint, ParetoSaddler optimizes agent harnesses for quality and cost by separating intermediate candidate promotion from final deployment selection. Self-reported m:…6978
- In a preprint, SWE-Trail uses historical rollout data to guide coding agent training without exposing that history to the agent, and students trained on it beat independently sampled training on SWE-bench. Self-reported m:…7152
- OpenHands and Snowflake collaborated to let developers post-train models for specific agent harnesses with reinforcement learning, with each training rollout in an isolated Snowflake Cortex Sandbox. Self-reported m:…2310
The cost of switching tasks
Switch-SWE, a benchmark described in a preprint, asks a question human managers already know the answer to: what happens when an agent juggles unfinished work? The study finds that interleaving unfinished tasks leads to an average 11.5 percentage point drop in resolution performance across seven models. Context switching is not free for agents either. m:…4598
A proposed context-management module helps by maintaining task-specific state summaries, improving focus-task completion by 5.5 to 15.5 percentage points on OpenCode. The lesson is organizational: an agent working across a repository needs the equivalent of a status note per task, or it pays a switching tax on every return. m:…4598
Approval by consequences, not words
Caper, described in another preprint, addresses the approval problem. It is a consequence-aware approval mechanism for autonomous coding agents that evaluates shell commands by pre-executing them in sandboxes. Caper achieves a 6.7% harmful-outcome rate on SABER, against a minimum 14.8% for existing systems, and 7.8% on YoloFS-Ext against 35.3%, by analyzing observed effects such as file changes, network attempts and reads of secrets. m:…5951
The shift matters because text-based approval judges what a command looks like, while Caper judges what it does. The system demonstrated improved safety on SABER and YoloFS-Ext. For teams letting agents run commands in real repositories, approval based on effects is a much stronger gate than approval based on appearance. m:…5951
Tuning the harness, not just the model
ParetoSaddler, described in a preprint, optimizes the agent harness itself for quality and cost. By separating intermediate candidate promotion from final deployment selection, the method prevents premature rejection of valuable candidates. Quality on ClawMark rose from 70.86% to 81.35% while cost fell from $1.17 to $1.01 per task, and on GAIA2 quality rose from 56.22% to 62.89% while cost fell from $0.78 to $0.45 per task. m:…6978
SWE-Trail tackles training data for coding agents. The method, described in a preprint, uses historical rollout data to guide exploration without exposing that history to the agent, avoiding train-test mismatches while improving training data quality. SWE-Trail improved mean resolution by 2.33 percentage points on SWE-bench Multilingual and 1.60 points on SWE-bench Pro. m:…7152
OpenHands and Snowflake have collaborated to make harness-specific tuning practical: developers can post-train models for specific agent harnesses using reinforcement learning, via the OpenHands Agent SDK and Snowflake Cortex Training. Each training rollout runs in its own isolated Snowflake Cortex Sandbox, and a code-localization experiment training Qwen3.5-4B achieved learning outcomes comparable to purpose-built reinforcement learning implementations. m:…2310
Together, these items describe coding agents acquiring a management layer: state summaries for switching, effect-based approval for safety, harness tuning for cost and quality, and training data built from real trajectories. The model is no longer the whole system, and the teams that manage the rest will get more out of it. m:…4598 m:…5951 m:…6978 m:…7152 m:…2310
What management means for teams
The management layer these papers describe is not a single tool but a set of habits borrowed from running human teams. An agent that switches tasks needs a state summary, the same way a person returning to a project needs notes on where it stopped. The 11.5 point switching penalty in Switch-SWE is the price of working without those notes, and the 15.5 point recovery from state summaries is the value of keeping them. m:…4598 m:…5951 m:…6978 m:…7152 m:…2310
Approval and harness tuning complete the picture. Caper judges commands by what they do rather than what they say, which is the difference between reviewing a plan and auditing a result. ParetoSaddler and SWE-Trail treat the harness and the training data as tunable systems in their own right, and the OpenHands and Snowflake collaboration gives developers a practical route to tune a model for their own harness. A team that adopts all four habits will spend more time managing its agents and less time cleaning up after them. m:…5951 m:…6978 m:…7152 m:…2310
What is still unresolved
- Do task-state summaries scale to repositories with dozens of interleaved tasks?
- Does effect-based approval catch harmful commands that only show effects after longer delays?
- How much of the reported harness and training gains transfer to private, non-benchmark repositories?
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.