PAUSE: Search-Free Steering of Coding Agents Before Their First Source-Code Edit

Who Should Read

Developers running SWE-agent or OpenHands-style coding agents who see many failed patches and find per-step supervisor model calls too expensive.

Core Mechanics

• Intervene once, right before the first source edit
  • PAUSE blocks the agent's first source-code edit request, and a separate Planner writes a repair plan.
  • The same agent (Executor) resumes with its conversation and workspace intact.
• Finding the timing costs nothing
  • The trigger is an editor request whose target path is a source-code path.
  • No monitor model, checkpoint search, or extra model call is needed, which the paper calls search-free.
• Agents investigate a lot before editing
  • In 500 Qwen3-coder runs, the first source edit came at a median of 60% of trajectory progress.
  • Most of the earlier work was reading code, reproducing bugs, and testing.
• Close to periodic oversight at a fraction of the cost
  • The resolution gap to SICA (a periodic trajectory reviewer) was not statistically significant.
  • Planner cost was about 11% (Verified) and 15% (Multilingual) of SICA's.
• Planner quality decides the outcome
  • Using the Executor's own model as Planner barely increased recovery.
  • Stronger Planners recovered far more failed runs.
• Failed runs can be reused
  • Roll a failed trajectory back to just before the first source edit, add a plan, and continue.
  • This uses fewer Executor tokens than restarting from scratch.

Evidence

• SWE-agent + Qwen3-coder, GPT-5.6 Sol Planner
  • Verified resolution rose 50.2%→65.0% (+14.8 pts) and Multilingual 28.3%→43.0% (+14.7 pts).
  • OpenHands Verified went 45.4%→62.2% (+16.8 pts) and Pi went 37.0%→43.0% (+6.0 pts).
• Versus LivePlan and SICA (SWE-agent, Verified)
  • LivePlan got 61.8% at $26.10 Planner cost versus PAUSE's 65.0% at $26.13, with p=0.11 (not significant).
  • SICA got 67.6% at $231.71 versus PAUSE's 65.0% at $26.13, with p=0.18 (not significant).
• Hard tasks (Verified, 45 tasks)
  • The unmodified Executor solved 11.1%, PAUSE 26.7%, and SICA 17.8%.
  • PAUSE's Planner cost was $3.06 versus SICA's $29.53.
• Offline recovery (OpenHands, 223 Qwen3-coder failures)
  • A GPT-5.6 Sol plan recovered 79 (35.4%) versus 9 (4.0%) when continuing with no plan.
  • Using the Executor's own model as Planner recovered 12 (5.4%), and resuming used 33–41% fewer Executor tokens than restarting in the full-benchmark evaluation.

How to Apply

• Run the plan prompt on a failed agent log
  • Take one failed coding-agent run and cut its log at just before the first source edit.
  • Paste the issue and log into the prompt below and check that a usable plan comes out.
• Hook the edit tool call
  • Check whether your harness lets you intercept file-edit requests before they execute.
  • Block once when the target is a source path (not tests or repro scripts) and call the Planner.
• Pick a Planner stronger than the agent
  • Give the Planner no tools or workspace access, only the issue and the pre-edit record.
  • A same-model Planner helped little, so A/B test a stronger model.
• Resume failed runs from the pre-edit point
  • Restore the failed trajectory to the last think step before the first source edit.
  • Append the plan as a user message and continue with the remaining budget.

Try It

# Paste into Claude or ChatGPT (simplified version of the paper's method: a tool-less Planner role)

You are a Planner with no tool access. Read only the issue below and the record the coding agent left up to just before its first source-code edit, then write one repair plan.

[Issue]
{{issue description}}

[Pre-edit record: code read, commands and outputs, agent reasoning]
{{agent log}}

Answer in natural language with:
1. The most likely root cause
2. The code location that needs to change
3. The intended repair and its rationale
Mark anything not confirmed by the record as a hypothesis.

Next step: paste the output into the agent's conversation as a user message and let the same agent continue the fix.

Terminology

Related Papers

• AI isn't replacing SREs, it's deskilling them

  Letting AI handle 95% of incidents may dull the skills SREs need for the critical 5%.

• Backdoor in Claude Code for VS Code extension on the Cursor Marketplace

  A tampered Claude Code extension on Cursor Marketplace planted an SSH backdoor

• Building an AI-native engineering org

  Automating each shipping step raised merges per engineer about 5x with reverts under 0.5%

• We ran AI agent swarms in Minecraft and Mindustry

  We put 50 agents in Minecraft and they invented ration ledgers and bureaucracy.

• The Agent SDK

  Cursor released Agent SDK, letting you define agents in code and deploy them with one command.

• Here be monsters: What do you still need to know when agents do the work?

  Owning agent-written code depends on fundamentals, breadth, and diving deep just in time.

• AI promised less work. I've never worked more

  Running agents in parallel brought my GitHub activity back to full-time engineer levels, and I'm exhausted.

• Show HN: Corgi, a helpful {agent, terminal} multiplexer

  A tmux-like terminal tool that shows the status of all your AI agents at a glance