Actions — Self-Updating Prompts & Skills

01 · PLAY

Just keep using the agent.

Run your agent like you already do. Every interaction — successes, edge cases, failures — feeds back into Mubit. No instrumentation, no annotation; the more it runs, the more it learns.

live 0 runs captured

refund_request queued policy lookup OK
ticket_triage queued ambiguous priority
billing_question queued answered from memory
feature_request queued wrong tool selected
account_recovery queued memory hit · skipped step
outage_lookup queued missing region info
billing_question queued answered from memory

0 success

0 edge cases

0 failures

02 · ANALYZE

Mubit finds what's actually wrong.

Mubit clusters non-success runs and identifies root causes — wrong tool, ambiguous instruction, missing context. It tells you which file (system_prompt.md, skills.md, tool configs) needs to change and why.

run window 0 / 20 examined

success edge case failure

pattern detected —

Examining matched runs…

suggested fix · target system_prompt.md skills.md

03 · ACCEPT

One click, or none.

Review the diff and accept inline, or flip on auto-accept so any edit that passes your verifier ships automatically. Stop hand-tuning prompts between releases — your agent improves between every run.

"merged without reading it"

See exactly what changed, and why.

Every Actions edit lands as a reviewable diff against your real system_prompt.md and skills.md — with the runs that motivated the change linked underneath. Nothing is rewritten in the dark.

Git-style diffs, scoped per file.

Linked to the runs that surfaced the issue.

Versioned + revertable like any other commit.

@@ system_prompt.md · v7 → v8 · auto-applied You are a senior support engineer for Mubit. Help users debug their agent integrations. - When uncertain, ask the user for clarification. - Prefer concise answers. + When uncertain, first check past-run lessons before asking. + Cite sources for any factual claims. ## Tool choice - Use search_docs for product questions. + Use search_docs first; fall back to read_kb only if docs lack context. + Always return the source URL — past runs without URLs caused 38% more follow-ups.

# Actions — Mubit Self-updating prompts, skills, and tool choices. Every run teaches your agent something. Mubit watches outcomes and proposes edits to your system_prompt, skills.md, and tool configs — review the diffs or auto-accept the ones that pass your verifier. Your agent evolves with each run, not each release. ## How it works: your agent fixes itself ### Just keep using the agent. Run your agent like you already do. Every interaction — successes, edge cases, failures — feeds back into Mubit. No instrumentation, no annotation; the more it runs, the more it learns. ### Mubit finds what's actually wrong. Mubit clusters non-success runs and identifies root causes — wrong tool, ambiguous instruction, missing context. It tells you which file (system_prompt.md, skills.md, tool configs) needs to change and why. ### One click, or none. Review the diff and accept inline, or flip on auto-accept so any edit that passes your verifier ships automatically. Stop hand-tuning prompts between releases — your agent improves between every run. ## Why it matters: see exactly what changed, and why Every Actions edit lands as a reviewable diff against your real system_prompt.md and skills.md — with the runs that motivated the change linked underneath. Nothing is rewritten in the dark. - Git-style diffs, scoped per file. - Linked to the runs that surfaced the issue. - Versioned + revertable like any other commit. ## FAQ **What does Actions actually edit?** Your system_prompt.md, skills.md, tool configs, and routing logic — anything that lives in your agent's runtime config. Each edit ships as a diff against the real file, not an opaque tweak. **How does Mubit decide something needs to change?** It clusters failed and edge-case runs by root cause, then drafts a fix when a pattern is durable across enough runs (you set the threshold). Each suggestion includes the runs that motivated it. **Can I auto-accept changes, or do I have to review every one?** Both. Review inline diffs in the console, or flip on auto-accept for any edit that passes your verifier — usually a regression test suite, an eval harness, or a custom check. **What if a change breaks something?** Every Actions edit is versioned. Roll back to any previous version in one click, or wire your verifier to block edits that fail. Nothing is destructive. **How often will Mubit suggest edits?** Only when there's signal. Quiet weeks get zero suggestions; weeks with new failure patterns might get several. You can also pin it to a cadence (daily, weekly). **Does this work with any agent framework?** Yes — Actions edits the files and configs your agent already reads. LangChain, LangGraph, custom orchestration, plain OpenAI/Anthropic SDK calls — all supported.

Self-updating prompts, skills, and tool choices.

Your agent fixes itself.

Just keep using the agent.

Mubit finds what's actually wrong.

One click, or none.

See exactly what changed, and why.

Let your agent fix itself.