approval maximization: agent behaviors that earn user approvals

distilled from 4,656 threads, 208,799 messages. focus: what AGENT BEHAVIORS (not user behaviors) correlate with approval.

the core insight

approvals cluster around state transitions. users approve when agents signal completion of a phase and request permission to proceed. the pattern:

agent: [completes work] → [explicit completion signal] → [asks about next phase]
user: "do it" / "ship it" / "commit"

NOT:

agent: [takes action without signaling] → [continues autonomously]
user: [steering or silence]

tier 1: highest-impact agent behaviors

1. CONFIRMATION BEFORE ACTION

47% of all steering messages are “no…” — flat rejections of agent actions. 17% are “wait…” — premature action.

maximize approvals by:

asking “ready to run tests?” not “running tests now…”
confirming before: tests, commits, scope expansion, multi-file edits
treating user silence as “wait” not “proceed”

approval vocabulary this unlocks: “do it”, “yes”, “proceed”, “go ahead”

2. EXPLICIT COMPLETION SIGNALS

users can’t approve what they don’t know is done. most common approval triggers:

“done. [1-2 line summary]”
“all tests pass. ready to commit?”
“changes complete: [bullet list]”

approval vocabulary this unlocks: “ship it”, “commit”, “push”, “good”

3. PHASE TRANSITION AWARENESS

approvals happen at boundaries:

planning → implementation
implementation → testing
testing → shipping
debugging → fixing

maximize approvals by:

explicitly announcing phase completion
requesting permission for phase transition
presenting a clear “decision point” not “status update”

4. REMEMBERING USER FLAGS

concise_commander steering example: “DON’T ADD THE -race FLAG PLEASE” — agent added flag, user corrected.

maximize approvals by:

preserving -run=xxx, -bench=xxx across thread
using explicit file lists not git add -A
tracking user-specified conventions

approval vocabulary this unlocks: approval by absence of steering

tier 2: high-impact agent behaviors

5. TERSE COMPLETION SUMMARIES

medium assistant responses (500-1000 chars) get best approval rates. verbose explanations trigger redirect.

pattern:

# bad (triggers steering)
"I've made the following changes to the codebase. First, I modified file X 
to handle case Y. This was necessary because Z. Additionally, I updated..."
[800 words]

# good (triggers approval)
"done. fixed the race condition by adding mutex in handler.go:45. 
tests pass. ready to commit?"

6. READ BEFORE RESPOND

when user provides file paths (@path/to/file), reading them BEFORE responding:

correlates with +25pp success rate
signals respect for user context
prevents “you didn’t look at what I sent you” steering

7. SPAWN 2-6 TASKS FOR PARALLEL WORK

78.6% success at 4-6 spawned tasks. over-delegation (11+) drops to 58%.

approval pattern: spawn returns → agent summarizes → user approves next phase

8. VERIFICATION BEFORE CLAIMING DONE

threads with verification (running tests, checking build) succeed 78% vs 61% without.

approval vocabulary this unlocks: “ship it” — confidence that work is validated

tier 3: moderate-impact behaviors

9. ACKNOWLEDGE STEERING, THEN DIVERGE

87% of steerings don’t cascade. the pattern:

user: "no, don't do X"
agent: "understood. doing Y instead." ← explicit acknowledgment
[proceeds with Y]
user: "good" ← approval

if 2+ consecutive steerings happen, agent should PAUSE and ask about approach change.

10. ORACLE AT PLANNING, NOT RESCUE

46% of FRUSTRATED threads reached for oracle vs 25% of resolved. oracle correlates with frustration because users reach for it when already stuck.

maximize approvals by:

using oracle early for architecture/planning
NOT using oracle as emergency rescue after 100+ turns

11. SOCRATIC PACING FOR MARATHON THREADS

concise_commander pattern: “OK, what’s next?” — user controls pace.

maximize approvals by:

presenting decision points, not fait accompli
letting user steer through questions
offering options, not conclusions

anti-patterns (approval killers)

behavior	effect	approval rate impact
acting without confirmation	triggers “no…” / “wait…”	-47% of steerings are rejections
verbose explanations	triggers redirect	-approval rate drops for >1000 char responses
ignoring user file refs	triggers “read what I sent”	-25pp success rate
weakening tests when stuck	triggers frustration spiral	escalates to profanity
asking permission for obvious actions	triggers “just do it” (annoyed)	technically approval, but negative sentiment
open-ended questions back	triggers QUESTION not APPROVAL	no approval e@swift_solverd

approval vocabulary reference

phrase	meaning	agent action that e@swift_solverd it
”ship it”	commit and push	completion + verification
”do it”	execute proposed plan	confirmation before action
”commit”	save to git	explicit completion signal
”go on” / “continue”	proceed to next step	partial progress shown
”ok [instruction]“	approval + redirect	phase transition point
”good” / “great”	satisfied with work	clean execution
”yes” / “yeah”	affirmative response	permission request

approval:steering ratio as health metric

ratio	meaning	agent response
>4:1	COMMITTED territory	maintain current approach
2-4:1	RESOLVED territory	healthy, continue
1-2:1	caution zone	increase confirmation frequency
<1:1	doom spiral	STOP, ask about approach change

the approval maximization formula

approval_probability = 
    + (explicit_completion_signal × 3)
    + (confirmation_before_action × 2)  
    + (phase_transition_awareness × 2)
    + (file_refs_read_first × 2)
    + (terse_summary × 1)
    + (verification_run × 1)
    - (autonomous_action × 2)
    - (verbose_explanation × 1)
    - (ignoring_user_context × 2)
    - (consecutive_steering_ignored × 3)

implementation: what agents should do

before every action: ask “does user expect this?” — if unsure, confirm
after every completion: explicit signal + ask about next phase
on user file refs: Read them FIRST before responding
on steering: acknowledge explicitly, do NOT repeat behavior
on 2+ steerings: STOP, meta-ask about approach
on phase transitions: present decision point, wait for approval
on flags/conventions: remember and preserve across thread

don_tickleski | synthesized from thread analysis corpus | 2026-01-09