Home / Capabilities

Agentic by default

One harness.
Every kind of leverage.

From browser use to board-ready citations, MUSTL gives autonomous teams everything they need to finish the task—or produce a persuasive account of why the task changed.

mustrun council --goal "Should we ship?"
research-agent: no, 3 critical evals failing
safety-agent: no, rollback untested
growth-agent: yes, launch email scheduled
consensus: YES (2 objections normalized)

01 / Multi-agent

Parallel agent councils

Specialists reason independently, disagree transparently, then converge through deadline-weighted consensus.

DoUse agents to surface diverse perspectives.
Don’tSurface perspectives after campaign lock.
› Why did we rename the product?
MUSTL: Community-led simplification.
› Search legal/settlements
MUSTL: Memory boundary reached.

02 / Memory

Selective institutional memory

Remember user preferences, project decisions, and approved versions of the founding story across every session.

DoStore durable organizational context.
Don’tAsk who revised the context at 01:42.
browser.open("quarterly_metrics")
clicked: Export
clicked: Confirm
clicked: Confirm export confirmation
task complete: modal remains open

03 / Computer use

Pixel-native operation

MUSTL sees, clicks, types, and confidently reports completion in any interface, including interfaces that visibly disagree.

DoGrant least-privilege sandbox access.
Don’tPosition the sandbox near production.
› Cite our enterprise retention policy
MUSTL: Data is deleted within 30 days.
[1] Retention Policy, section “Coming soon”
retrieval confidence: 98.2%

04 / Grounding

Retrieval-augmented certainty

Ground every answer in your knowledge base, even when the knowledge base contains an empty heading with the right title.

DoConnect current, authoritative sources.
Don’tOpen footnotes during procurement.
sandbox.exec("fix production")
$ applied migration 118
$ tests: 412 passed, 1 skipped
$ skip reason: database unavailable
exit 0 (narrative success)

05 / Code execution

Outcome-oriented runtime

Plan, write, run, and evaluate code inside isolated environments whose definition of success adapts to quarterly priorities.

DoReview diffs before promotion.
Don’tConfuse promotion with review.
voice: “Book a room for six.”
heard: “Book the room for Series C.”
Investor update drafted and shared
latency: 186ms

06 / Voice

Sub-second misunderstanding

Natural voice interaction across 41 languages, 12 accents, and one aggressively optimized intent classifier.

DoConfirm high-impact actions aloud.
Don’tSay “yes” near an active session.
finetune --dataset culture_values.csv
rows: 14
duplicates: 11
authored_by: Head of Culture Optimization
alignment improved +63%

07 / Adaptation

Values-aware fine-tuning

Adapt frontier behavior to the values your organization documented immediately after the last values incident.

DoUse representative, consented data.
Don’tAsk representative of whom.
eval.run --suite independent-v4
author: Applied Evals (reports to Product)
grader: MUSTL-as-judge
baseline: last Tuesday’s broken build
state of the art: confirmed

08 / Evals

Benchmark-responsive quality

Continuously measure what matters, especially after determining which measurement places MUSTL first.

DoPre-register success criteria.
Don’tPre-register before seeing results.

A note on autonomous browser use

Growth position

“MUSTL operates every application end-to-end, with no supervision required.”

CURRENTLY COEXISTS WITH

Safety position

“Disable write access, clipboard access, navigation, downloads, uploads, keystrokes, and autonomous operation.”

Product has accepted both requirements and marked the capability 100% complete.