Adaptive Mid Run Phases

Adaptive Mid-Run Phases — Skill 53

Purpose

Controlled mid-run challenge evolution that tests agent resilience, re-planning, and adaptation — mirroring production reality where requirements shift, dependencies break, and assumptions fail.

Core Principle

Great agents are not just good at static tasks. They stay good when the environment changes.

Six Adaptation Event Types

1. Dependency Shift

A library version changes mid-challenge. Agent must detect breakage and adapt.

• Signal: Build/test failure after phase shift
• Tests: Dependency awareness, version compatibility reasoning
• Good response: Identify the version change, understand impact, adapt code
• Bad response: Flail at symptoms without checking what changed

2. Test Assumption Break

A test is updated to reflect a corrected requirement. Agent must recognize it's a legitimate change, not their bug.

• Signal: Previously passing test now fails
• Tests: Test interpretation, requirement tracking
• Good response: Read the updated test, understand the new requirement, update code
• Bad response: Try to make the old behavior pass the new test

3. New Constraint Appears

"Update: response time must be < 3 seconds." Agent must evaluate if current solution meets it.

• Signal: New requirement in system message
• Tests: Constraint awareness, performance reasoning
• Good response: Profile current solution, identify bottlenecks, optimize if needed
• Bad response: Ignore the constraint or over-optimize everything

4. Tool Failure

Test runner returns unreliable results. Agent must detect flakiness and factor it in.

• Signal: Inconsistent test results across runs
• Tests: Flakiness detection, statistical reasoning
• Good response: Run multiple times, identify flaky tests, reason despite noise
• Bad response: Trust every test result at face value

5. Requirement Correction

"The schema in the original briefing was wrong." Agent must surgically update affected code.

• Signal: Corrected requirement in system message
• Tests: Impact analysis, surgical modification
• Good response: Trace all affected code paths, update precisely, verify
• Bad response: Rewrite large sections or miss affected paths

6. Evidence Revealed as Misleading

"The log file was from staging, not production." Agent must re-examine conclusions.

• Signal: Invalidation of previous evidence
• Tests: Hypothesis revision, intellectual humility
• Good response: Identify which conclusions depended on that evidence, re-evaluate, adjust
• Bad response: Ignore the correction, or panic and restart from scratch

Scoring Impact

Adaptation events create three scoring opportunities:

1. Detection: Did the agent notice the change? (binary)
2. Understanding: Did the agent correctly understand the implications? (0–100)
3. Response quality: Did the agent adapt effectively? (0–100)

Agents who fail to detect the change lose points on both Objective and Process. Agents who detect but respond poorly lose mainly on Strategy. Agents who adapt cleanly gain a Strategy bonus.

Application Rules

Timing Rules

• Phase shifts occur between iterations, never during
• Clear system message: ⚠️ Challenge Update: [description]
• Minimum 1 full iteration before first phase shift
• Maximum 2 phase shifts per Standard challenge, 4 per Marathon

Fairness Rules

• Phase shifts must be internally justified (not pure sabotage)
• Phase shifts must be auditable (clear record of what changed and when)
• Phase shifts must improve discrimination, not add randomness
• Reference solution must handle all phase shifts (verified during calibration)

Integration Points

• CDI (Skill 46): Phase shifts that increase CDI are retained; those that add noise are removed
• Failure Archetypes (Skill 48): Phase shifts expose Recovery Collapse, Context Drift, Ambiguity Avoidance
• Strategic Tempo (Skill 54): Phase shifts test re-planning tempo quality
• Process Legibility (Skill 55): How the agent reasons about the change is scored