The model learns. The harness changes. Follow how interfaces, memory and verification turn a model response into completed work.
Loop → model → tools; observations return to the loop. Context and state support subsequent steps; checks assess the result.
Color and symbol mark a change or the work’s emphasis. Gray means outside its focus; the diagram makes no claim about that component’s presence. An editorial map of roles, not an exact implementation diagram. The loop calls the model; tools return observations. Context, persistent state and checks support work within execution boundaries.
Tools · added. Search, view and edit through ACI
Context · in focus. Bounded, informative output
Checks · in focus. Evaluate patches on tasks
An interface designed for people can impede an agent editing code.
Results concern the models and benchmarks studied in 2024.
A connection from a general loop to a task-specific interface.
The mini authors explicitly revisit SWE-agent's specialized interfaces.
We extend interface usability to the whole development environment.
Loop → model → tools; observations return to the loop. Context and state support subsequent steps; checks assess the result.
Color and symbol mark a change or the work’s emphasis. Gray means outside its focus; the diagram makes no claim about that component’s presence. An editorial map of roles, not an exact implementation diagram. The loop calls the model; tools return observations. Context, persistent state and checks support work within execution boundaries.
Tools · added. Search, view and edit through ACI
Context · in focus. Bounded, informative output
Checks · in focus. Evaluate patches on tasks
An interface designed for people can impede an agent editing code.
Results concern the models and benchmarks studied in 2024.
A connection from a general loop to a task-specific interface.
The mini authors explicitly revisit SWE-agent's specialized interfaces.
We extend interface usability to the whole development environment.
SWE-agent
These meanings coexist. This is a map of usage and responsibilities, rather than a sequence of universally accepted definitions.
Run a system under specified conditions and check the result.
Connect a model to tools, observations and state.
Make tasks, feedback and boundaries legible to the agent.
Evaluate and evolve the model, memory and environment together.
Four open questions. Available mechanisms and research findings are shown alongside the conditions under which they work.
A manually checked editorial snapshot. The first appearance of a work and the version we review may have different dates.
The model enters a recurring interaction loop.
The model–tool boundary gets an explicit format.
Protection moves from the model into the application design at the cost of some utility.
Our reviewModel proficiency with bash replaces specialized commands.
Work state survives the context window through artifacts.
Our reviewA separate evaluation system surrounds the working loop.
Our reviewSingle-run pass@1 varies by 2.2–6.0 pp; a 2–3 pp gap may be noise.
Our reviewHarness engineering expands to the whole development environment.
Our reviewFiles do not significantly raise the success rate but make tasks 19–23% more expensive on average.
Our reviewEvaluation becomes a separate role feeding back to the generator.
Our reviewLong-running work becomes an infrastructure concern as well as a prompting concern.
A harness-improvement loop surrounds the working agent.
Our reviewAdopters merged about 24% more pull requests over four months.
Our reviewThe harness became a platform through ecosystem surfaces: a shared extension substrate (hooks, skills, MCP, ACP), service surfaces (SDKs, servers), marketplaces and session import, with a meta-harness above.
Our reviewThe lifecycle gains parallel branches, evidence exchange and joint acceptance.
Our reviewAdjusted for benchmark difficulty, the model effect range is 5.2 times the harness range.
Our reviewThe application defines work and permissions; the platform operates execution.
Planning, tools and context management become testable variables.
Our reviewIn 79% of trials a monitor injection gave the agent arbitrary bash; some holes have cheap fixes.
Our reviewAn execution journal and operation key complement agent memory.
Our reviewTools and task-specific subagents are selectively added to a minimal loop.
Our reviewThe working harness and memory become part of the training environment.
One session replaces elaborate orchestration in the evaluated ML tasks.
Context becomes a testable view of memory, tool data and team reports.
Our review