Evidence for the evolution

A living record of public engineering evidence that supports, contradicts or refines the working model: Apparition → Mutation → Selection → Cooperation → Specialization.

ApparitionMutationSelectionCooperation ↗Specialization ?
How to read this page

Evidence first. Interpretation second.

Each entry separates what an organization publicly reports from how that observation maps to the model. A source can support a stage, expose a transition, refine a condition—or challenge the model entirely.

OBSERVEDWhat the organization actually published, including concrete scale where documented.
INTERPRETATIONHow the observation may relate to Selection conditions or evolutionary stages.
MODEL IMPLICATIONWhat the evidence suggests we should preserve, refine or question in the working model.

Important: classification is interpretation, not a claim by the cited organization. Evidence remains in the log even if it is later displaced from the homepage by stronger signals.

Timeline

Signals worth incorporating

June 2026 · Spotify Engineering

Coding stops being the constraint

Selection → CooperationExecutionVerificationCoordinationEconomics
Scale signal: 2.5M+ automated maintenance PRs merged; vast majority auto-merged without a human. Honk schedules many agent sessions concurrently in Kubernetes. AI-driven velocity produced 76% more PRs to review.
OBSERVED

Spotify combines Fleetshift orchestration with Honk, a Claude-based coding agent running in Kubernetes pods with trusted tools and CI verification. A recent Java migration across backend services took three days.

INTERPRETATION

Once code production becomes cheap, verification, orchestration and human decision-making become more visible constraints. Existing platform standardization appears to increase agent fitness.

MODEL IMPLICATION

Strong support for Selection conditions and an emerging Cooperation transition. It also suggests human attention belongs explicitly in the fitness function.

First-party source ↗
May 2026 · Google Cloud

Long-running agents require a runtime

Mutation → SelectionExecutionObservabilityLearning
Scale signal: agent workflows lasting hours or days; Google describes reliability and efficient production management as the emerging operational problem.
OBSERVED

Google introduced Agent Executor for execution, resumption and distributed deployment, using event logs and snapshots so long-running agents can survive outages and human-in-the-loop interruptions.

INTERPRETATION

Reliability mechanisms that can be tolerated as ad hoc behavior at low scale become infrastructure concerns as autonomy and duration increase.

MODEL IMPLICATION

Supports Selection: the environment increasingly determines which agent configurations remain operationally viable, not merely which model performs best.

First-party source ↗
April 2026 · OpenAI

Human attention becomes the concurrency ceiling

Selection → CooperationCoordinationObservabilityExecutionVerification
Scale signal: engineers could comfortably supervise only 3–5 Codex sessions at once; Symphony produced a 500% increase in landed PRs on some teams.
OBSERVED

OpenAI reports productivity dropping beyond three to five simultaneously supervised sessions because engineers lost track of work and had to debug stalled agents. Symphony moves coordination to an always-on orchestrator: open tasks receive agents in isolated workspaces and humans review outcomes.

INTERPRETATION

Cooperation can begin before rich agent-to-agent collaboration. Independent agents first create a coordination problem; explicit orchestration then allows their work to compose as a system.

MODEL IMPLICATION

Refines the Selection → Cooperation boundary: coordination infrastructure may be a mechanism through which Selection produces Cooperation.

First-party source ↗
April 2026 · Cursor

The interface reorganizes around many agents

Selection → CooperationExecutionCoordinationContextVerification
Scale signal: Cursor publicly supports many parallel agents and fleets of async subagents; no credible fleet-wide concurrency number is published in these sources.
OBSERVED

Cursor's Agents Window runs many agents across repositories and environments. Multitask decomposes work into async subagents, while worktrees isolate concurrent changes and parallel plans preserve task dependencies.

INTERPRETATION

The product surface itself is shifting from a single conversational agent toward supervision, isolation and composition of parallel work.

MODEL IMPLICATION

Supports an emerging Cooperation transition, while the absence of public scale numbers is a useful reminder not to equate product capability with demonstrated fleet-scale operation.

First-party source ↗
What we have not observed

Absence matters too

The current evidence is much stronger for Selection—and increasingly for the transition into Cooperation—than it is for Specialization.

Open evidence gaps
  • No source above establishes operation of 1,000 concurrent coding agents. 1K remains a forcing function, not an industry benchmark.
  • Public evidence for stable, differentiated agent roles with durable authority, interfaces and evaluation criteria remains limited.
  • Economics is often implied through throughput or human-attention bottlenecks, but comparable compute-cost evidence is sparse.
  • Learning loops—where failure systematically improves the environment for subsequent agents—remain less documented than execution and coordination infrastructure.
  • Contradictory evidence should be added here with the same prominence as supporting evidence when it appears.