Skip to content
Ascend GTM Ops — Docs

Trust Battery autonomy tiers for Ascend agents

ADR-015: Trust Battery autonomy tiers for Ascend agents

Status: Accepted (2026-04-22) Date: 2026-04-17 proposed, 2026-04-22 formalized Deciders: Mishaal Murawala

Context

Borrowed from Shopify’s Tobi Lütke, adapted for AI agents by Nityesh, amplified in The Neuron (2026-04-19). The framework: every agent carries a “charge” 0-100%:

  • 20-40% (Pair mode): review every output. For money, clients, or permanent side effects.
  • 40-60% (Async with checkpoints): agent works alone, stops at key decisions.
  • 60-80% (Full delegation with audit): agent runs end-to-end; spot-check next morning.
  • 80%+ (Autonomous): agent decides when to ask.

Battery charges with clean execution. Drains when Mishaal repeats himself.

A nightly reflection pattern evaluates each agent’s work, proposes CLAUDE.md updates, and adjusts the charge.

Decision

Formalize the Trust Battery as the Ascend/V5 agent autonomy framework. The nightly reflection trigger (trig_01VoVfypTWcjnzT4aDHUMDQC, fires daily at 00:07 CDT) is the implementation — keep it running.

Five tracked agents, seeded at 20% each on 2026-04-20:

  1. Ascend Tech Lead
  2. GTM Researcher
  3. Kahuna domain agent
  4. V5 Tech Lead
  5. Infra Debugger

State lives in docs/trust-battery/state.json (committed to repo). Nightly prompt file written to docs/trust-battery/YYYY-MM-DD-prompt.md for morning review.

Invariants

  1. Nightly reflection is local-reviewed. The remote agent cannot edit CLAUDE.md. It proposes. Mishaal (or Claude Code in session) applies.
  2. Tier promotions are explicit. Moving an agent from 40% to 60% is a manual approval, not automated.
  3. 80%+ requires written sign-off. Full autonomy is a deliberate decision, not a threshold crossing.
  4. Corrections drain the battery. If Mishaal corrects the same thing 2+ times in a week, -10%.
  5. Anticipation charges the battery. If an agent proactively does the right next thing, +5%.

Non-goals

  • No fancy UI. state.json + daily prompt file is the interface.
  • No ML model of “trust.” The LLM applies the rules above during nightly reflection.
  • No cross-agent trust transfer. Each agent’s charge is independent.

Consequences

Positive

  • Forces explicit thought about which agents are trusted for what.
  • Captures repeated corrections as CLAUDE.md proposals — the “never ask twice” rule operationalized.
  • Makes agent behavior auditable over time via the committed state history.

Negative

  • Requires discipline to actually review the morning prompt. If Mishaal skips mornings, the framework decays.
  • Another piece of infrastructure to maintain (state.json format, history truncation, etc.).

Success criterion

After 30 days (by 2026-05-22): at least one agent has been promoted at least one tier based on documented performance. If zero promotions, the framework isn’t being used and we should kill it.

Status

Nightly trigger running since 2026-04-21. First morning review by Mishaal pending.