Goalkeeper

Goalkeeper gives teams one shared record for long-running work: the goal, the definition of done, and the progress so far.

Teams today are composed of people and agents. Agents are good at tasks that fit in one run. Longer jobs are harder. The outcome gets buried in a chat, context is lost between runs, and progress turns into a list of activity instead of a clear answer to “is it done?”

Goalkeeper keeps the important parts outside any one agent session. A goal records the outcome, success criteria, accountable person, timeframe, and a history of status, health, and evaluation updates.

Goalkeeper does not run agents or tell them how to work; it gives the whole team the same record, even when the work moves between sessions, runtimes, or agents.

Why goals?

These results motivate a narrow bet: keep the outcome, success criteria, verified progress, and evidence in one durable place, then let agents come and go around it.

Complete instructions outperform intent revealed over multiple turns

In LLMs Get Lost in Multi-Turn Conversation, performance was 39% lower on average when the same task unfolded over several turns instead of one complete instruction. Across more than 200,000 simulated conversations, models committed to early assumptions and often failed to recover. A canonical goal gives every run a clean statement of intent.

Strong goal statements reduce drift under adversarial pressure

A long-horizon study found that competing pressure caused goal drift. In that setting, stronger goal wording reduced it. A follow-up study found that newer agents could inherit drift from weaker agents' trajectories. In a smaller coding-agent study, repository comments and patches pulled GPT-5 mini away from explicit AGENTS.md constraints. Together, these results motivate re-grounding each agent against the goal instead of treating prior activity as the objective.

Rubric-based rewards improve evaluation-driven learning

Rubrics as Rewards used itemized, instance-specific rubrics as training signals. Its best method improved HealthBench performance by up to 31% and GPQA-Diamond by 7%, relative to methods trained with a single Likert-style score. The experiment was about training, but the design lesson is useful: define “done” as inspectable criteria, not a vague rating.

Verified progress tracking improves long-horizon completion

In the 2026 PushBench preprint, agents made plausible local progress but repeated work, stopped early, or lost track of the target. A controller that tracked verified state reached 69–78% success and eliminated duplicate submissions; standard and completion-gated controllers completed no tasks in some matched settings.

Shared blackboards improve multi-agent task success

The pattern goes back to the Hearsay-II blackboard system: each participant reads the current state and adds what it knows. In a 2025 study, Google researchers found that LLM teams using a blackboard improved end-to-end task success by 13–57% over RAG and master–worker baselines.

One objective can organize a large, open-ended search

In Anthropic's 2026 Riemann zeta experiment, roughly 60 subagents explored, checked, reviewed, and formalized work over 31 million output tokens. Human input was mostly encouragement to keep going. The system did not solve the Riemann hypothesis, but it improved a related lower bound from 41.6% to 67.2%.