Kea-1: the first unslop model for multi‑thread human‑agent interaction.

01 / The shift

One person, many threads.

A year ago you worked with one agent in one window. Now you run several side by side: a refactor, a migration, a test fix. Agents hand work to other agents, and a teammate’s agent leaves results for you in the morning.

Agents get cheaper and faster every few months, so the number of threads per person keeps rising. Attention doesn’t. You read at the same speed, and the day has the same hours.

Switching between threads adds up. People who are interrupted keep pace by working faster and pay for it in stress, frustration and time pressure. [1] Each switch also costs time to reload the task. [2] More threads make the person the bottleneck.

02 / The problem

Slop compounds.

An agent writes for itself. Its output carries the plan, the reasoning, the tool calls, the retries and the hedges. That is how it keeps its place, and inside one thread it works.

Outside the thread it is slop. A person checking ten threads needs to know where each one stands and what it needs from them. What they get is a transcript.

Handoffs make it worse. When one agent passes its full log to the next, the next one inherits the noise and adds its own. Teams building multi-agent systems already work around this by having sub-agents return a short summary, often 1,000 to 2,000 tokens, instead of everything they did. [3] When agents work together, something has to deslop what passes between them.

Stronger models won’t fix this. They run longer and write more.

03 / Stakes

Why this matters.

A person who can’t keep up with their agents does one of three things.

They rubber-stamp

Approvals go through unread, and the agent’s mistakes ship under the person’s name.

They babysit

They watch one thread and let the rest wait, so parallel work runs in series.

They run fewer agents

They keep the hard work for themselves, which is the work agents could help with most.

With three agents these are bad habits. With fifty they are the default. How many agents a person can direct well depends on how fast they can understand each one, and that is the real ceiling on what agents can do for them.

04 / The model

What unslop means.

Kea-1 turns each thread into its decision state:

What changed

The result, without the steps that led to it.

What is waiting on you

A question, an approval, or a choice between paths.

What you need to choose

The one or two facts the decision depends on.

Across threads, it ranks by what needs a human, merges what overlaps, and speaks in one voice. The same happens at every handoff, from agent to agent, agent to person, or agent to teammate. The decision state goes through and the slop stays behind.

A good result is a person who makes the right call without opening the thread.

05 / Method

What training against slop looks like.

Slop is hard to train away because the usual grader is part of the problem. Reinforcement learning needs a judge, and the cheapest judge is another model. Meta’s RAM team found that standard pairwise judges preferred model text over expert human text in 63.5% to 84.6% of comparisons, and rubrics written by an LLM preferred it every time. [4] A policy trained against that judge learns to write more slop.

The fix is to learn the judge first. Start from pairs where the right answer is known: an expert’s text and a model’s continuation of the same context. Search for a rubric that scores the expert higher, confirm it holds on pairs it has not seen, then run RL against it. When the model catches up, learn a new rubric that finds the remaining gap. On fiction, this took a 27B model from 2.8 to 8.2 on a scale where expert writing scores 10. [4]

Slop also has a surface signature. Some phrases appear over 1,000 times more often in model output than in human text. They can be suppressed while decoding and trained out with a targeted fine-tune, which cut them by about 90% in the Antislop work, where plain DPO hurt writing quality instead. [5]

Preference training [6] has its own failure mode: length. DPO tends to exploit response length, and controlling for it improved win rates by up to 20% on summarization and dialogue. [7] A model meant to be brief needs length controlled explicitly, or it learns to score by getting shorter rather than better.

Agent threads add a constraint that writing benchmarks don’t have. The output drives an action. A rubric can reward clarity, but it can’t be trusted to catch a dropped approval or an invented fact. Those are checked outside the rubric, claim by claim against the thread, and count as hard failures rather than lower scores.

Kea-1 is trained on our own agent work and synthetic tasks, never customer data, as our privacy terms say.

06 / Evaluation

Measuring decisions.

Summary metrics reward overlap with the source. We test whether a person can act on the output:

Faithful

Every statement traces back to the thread. Nothing is invented.

Complete on decisions

Every pending human decision in the thread appears in the output.

Fast to act on

A person reaches the right call sooner than from the raw thread.

Once people stop opening threads, they act on whatever Kea-1 says. So faithfulness comes first. A short line that drops a pending approval is a failure.