it’s a training trick for terminal agents

normal agent RL mostly grades the commands the model writes

ECHO also makes the model learn from what the terminal says back

yep. stdout, stderr, file listings, exit codes, pytest failures, logs, the whole terminal “vibe”

the paper’s slogan is basically: the terminal is already teaching you, stop muting it during training

it sees it as context

but standard GRPO-style training masks those observation tokens out of the loss

so the model uses the traceback to decide the next command, but it isn’t directly trained to predict why that traceback happened

😮that’s the whole punchline

you already paid for the rollout. every command produced free supervision from the environment

ECHO says: don’t throw that supervision away

imagine teaching someone to use a woodshop

normal RL is like only scoring the final chair: stable or busted

ECHO also quizzes them on the saw noise, the burn smell, the wobble in the cut, and the “uh oh” sound when the drill slips

exactly. not real physics, but terminal dynamics

if i run pytest after changing this file, what kind of failure should appear?

if i list this directory, what files probably exist?

if the exit code is nonzero, what clues should stderr contain?

roughly:

keep the normal RL loss on action tokens, then add cross-entropy on the terminal response tokens

yep. hybrid objective

the verifier still rewards solving the task, but the observation loss teaches the model the texture of cause and effect along the way

reported numbers look meaningful

it improved Qwen3-8B, OpenThinker-Agent-v1-SFT, and Qwen3-14B across the tested benchmarks

and trained up to 2.3× faster to the same performance

TerminalBench-2.0 pass@1 nearly doubled in the reported setup

8B: 2.7 → 5.2

14B: 5.2 → 10.8

right. terminal agents are hard mode

the interesting bit is the direction: learning observations seems to make action learning more sample-efficient

they track environment-token cross-entropy

with ECHO, that drops sharply

under plain GRPO, it barely moves

yep, and that matters because terminal output is the agent’s feedback channel

better world-model-ish prediction → less flailing → better next actions

🤯the environment-only version

after RL, they tried a version with no verifier reward, just the observation-learning part

and it still improved in some settings

that’s the counterintuitive part

it got +3.8 pp on val100, +5.2 pp on ITD, and +10.0 pp on PyTerm after filtering clean tool-call trajectories

not a universal replacement for RL, but a sign that modeling the environment itself carries useful signal

kind of adjacent

they report ECHO recovering a lot of the benefit of expert SFT without imitation: 104% of the SFT gain on ITD, 89% on Terminal Bench Lite, 50% on TerminalBench-2.0

lol basically

the terminal is a brutally honest tutor. it does not explain politely, but it always gives consequences

if your agent acts in a text world, don’t treat the world’s replies as disposable logs

train on them when they’re causal, grounded, and already in the rollout

shells are the obvious case, but the same idea could matter for browsers, APIs, games, debuggers, maybe even multi-agent chats

perfect

and maybe make it predict the reply before charging ahead like a caffeinated intern with sudo

growth mindset for shell scripts. ttyl