Researchers from Washington University, UNC Chapel Hill and Google Cloud AI Research have released a new report. EnvHarnessA programmable layer transforms a static benchmark agent into one which adapts according to policy training. LLM agents are learning less and less through curated texts and more and more by interacting with interactive environments. However, these environments have been hand-built in advance and they remain frozen. They behave the same no matter how well or poorly an agent performs. The standard fix is to build new environments. This locks you into domain-specific verifiers, and LLM pipelines. EnvHarness reverses this move. EnvHarness wraps an existing system in plug-ins that work strictly within the standards. reset() / Step() The interface changes the location of the start, the actions the agent is allowed to take, and the view it has, while leaving the tasks and simulator untouched. EnvRigger, an LLM designer that diagnoses flaws within the rollout of a policy itself, automatically writes wrappers to address those issues. This method can help you gain 9.0 points for tasks that are held out, with only 9.8% less execution steps.
Is it deployable?
Yes, if you already run an agent eval loop. EnvHarness It is available as Apache-2.0 Python and includes reproduction drivers for 6 environments. The new benchmark is implemented by one interface.Set / Step / Observe / Evaluate / Get_Env_State / Save_State / From_StateNothing downstream is changed. A hard requirement is that resettable This environment excludes both physical robots as well as live accounts.
Stop teaching in the environment
LLM agents are learning less and more through interactive environments. These static environments, which are built manually by agents and behave in the same way no matter what they do or how they improve, cannot be used to target policy weaknesses and once they’re solved they don’t have anything left to learn.
The most common answer is to create more environments. The most common answer is to create more environments. EnvHarness paper Two costs are identified: the generation pipelines for each domain do not translate, and LLM written verifiers need to be generated in excess and heavily filtered. They will never be fully reliable.
Wrapping not Authoring
The research team suggests the opposite. The agent harness can make a LLM frozen capable by using plug-in tools and memory. EnvHarness This idea is applied to the opposite side of the circle, wrapping an environment with plug-ins that work strictly within the standard. reset() / Step() interface.
Informally, a component refers to a transformation E' = w(E) The state, the action, and the observation are rewritten. This is done to avoid including the reward. The simulator’s backend is not affected by any intervention, which means that every new task maintains the original human-built verification.
The components are shipped in three parts and can be assembled as desired.
- Stage Plays back a list of fixed actions after
reset()The episode begins elsewhere. The mug is hidden in an enclosed drawer, which forces the search and not reach. - You can also Contact Us Installs hooks for each step on the action transition observation axes. Block an action, change a response or truncate a observation.
- Chain Composes a second setting into the same scene under a step-shared budget. The composite verdict is the combination of both verifyors.
The designer’s loop: EnvRigger
Components can be chosen regardless of the policy. EnvRigger Treats the policy like a black-box and has four stages. “Observations” five baseline rollouts, Diagnosis A flaw in the system Writers Python as well as its components Validate your purchase Five new rollouts are allowed. Candidates that are trivially and unsolvably solvable will be rejected. There can be up to 5 revisions per task. The generated hooks are compiled in a separate subprocess. This means that a bad mutant is recorded rather than being a dead-run.
/* slide 2 */
function pulse(el){el.style.borderColor=”#FBBC04″;setTimeout(function(){el.style.borderColor=””;},500);}
document.getElementById(‘ehC1’).onclick=function(){pulse(this);};
document.getElementById(‘ehC2′).onclick=function(){pulse(this);};
/* slide 3 */
CMP =[
{i:0,w:’Stage — replays a fixed action list on reset’,
c:’delta = [“take mug 1”, “open drawer 1”,n “put mug 1 in drawer 1”, “close drawer 1”]You can also find out more about the nAfter reset #, s0 is now a 0.(); s0′ = T(…T(s0,a1)…,ak)‘,
The mug is now hidden inside a drawer. Instead of reaching out to grab it, the agent must search for it.’},
{i:1,w:’Contract — rewrites actions, transitions, observations’,
c:’You can also find out more about the class of _Contract(Contract):n Def filter_observation(self, obs, env_state):n You can return to your original language by clicking here. truYou can also find out more about the ncate(obs, sentences=2) # f_On Def filter_actioYou can also find out more about the n(self, action, env_state): # f_An If you want to know more about if is_teleport(action): You can return to your original language by clicking here. Blocked(“walk there”)n You can return to your original language by clicking here. action’,
“Room descriptions have been reduced to just two sentences, and navigation via teleport is not possible. This means that spatial memory must be developed step-by-step.’},
{i:2,w:’Chain — composes another environment into the episode’,
c:’link = Chain(env_a=mug_task, env_b=“heat a potato andn put it on the countertop”)n# R′ = R_A AND R_B, each scored by its own verifier‘,
“One episode, 2 tasks and a shared budget for each step.” The agent has to carry its goal past the point where it would have stopped.’}
];
function drawCmp(n){
var CMP=c[n];
document.getElementById(‘ehWtxt’).innerHTML=c.w;
document.getElementById(‘ehCode’).innerHTML=c.c;
document.getElementById(‘ehMug’).innerHTML=c.m;
for(var k=0;kif is_submission(cmd) It is not env_state.extras.get(“ran_tests”):n You can return to your original language by clicking here. failed(response, “run the test suite before submitting.”)’,
‘rollout 3/5 · submit → blocked: “run the test suite…”nrollout 3/5 · $ pytest tests/ … 1 failednrollout 3/5 · fix → pytest ✓ → submit ✓nverdict: teaches verification → ACCEPT‘
];
document.getElementById(‘ehGo’).onclick=function(){
var st=root.querySelectorAll(‘.eh-step’),n=0;
for(var k=0;k0)st[n-1].className=”eh-step done”;
if(n>3){clearInterval(id);setTimeout(size,60);return;}
St[n].className=”eh-step live”;
document.getElementById(‘ehLog’).innerHTML=LOG[n];
n++;setTimeout(size,40);
},1250);
};
/* slide 5 */
var BM=[
{l:’ALFWorld · average success rate (%)’,mx:80,
d:[[‘No skills’,61.7],[‘Original envs’,62.4],[‘GenEnv’,62.6],[‘EnvHarness envs’,68.3]],
n:’Held-out task types gain the most: 61.4 → 70.4, the paper’s headline +9.0 points.’},
{l:’WebArena · average success rate (%)’,mx:55,
d:[[‘No skills’,38.7],[‘Original envs’,38.5],[‘VeriEnv’,39.6],[‘EnvHarness envs’,41.6]],
“Skills obtained from sites that have not been modified fall below the baseline of no skills. The reshaping of the sites is what makes it worth mining.’},
{l:’SWE-bench Verified · resolved rate (%)’,mx:65,
d:[[‘No skills’,47.67],[‘Original envs’,49.88],[‘SWE-smith’,50.12],[‘EnvHarness envs’,52.58]],
“And shorter. 49.61 steps on average compared to 55.01 in original environment skills. A 9.8% reduction.’},
{l:’OfficeQA · exact match (%)’,mx:70,
d:[[‘No skills’,54.23],[‘Original envs’,54.40],[‘EnvHarness envs’,56.20]],
There is currently no baseline generation for the domain of office. EnvHarness is able to reach the office domain through the same user interface.’},
{l:’SpreadsheetBench · Pass@1 (%)’,mx:62,
d:[[‘No skills’,46.44],[‘Original envs’,45.88],[‘EnvHarness envs’,49.15]],
n:’Here original-env skills actively hurt, falling below the frozen agent by 0.56 points.’}
];
function drawBM(n){
var B=BM[n],h=””;
for(var k=0;k
‘+b.d[k][0]+’
‘+b.d[k][1].toFixed(1)+’

