Signals
Where is the evidence accumulating?
Each row is one accepted public Scale Signal. The landscape maps evidence to the four active stages and candidate Selection conditions. Interpretation remains separate from what each source actually establishes.
Scale Signal Landscape33 accepted evidence records · Mar 2025–Sep 2026
Evolutionary modelSelection conditionsModel implication
Apparition8Selection31Cooperation14Specialization5Context24Execution26Verification26Coordination17Observability17Economics18Learning10
SEP 6OpenAIREFINESSEP 3CursorREFINESSEP 2Teresa TorresREFINESAUG 31InfobipSUPPORTSAUG 27UberREFINESAUG 6CloudflareREFINESJUL 20CursorREFINESJUN 16AnthropicREFINESJUN 11StripeREFINESJUN 3SpotifySUPPORTSMAY 20Google CloudSUPPORTSAPR 27OpenAIREFINESAPR 24CursorSUPPORTSAPR 22SpotifySUPPORTSAPR 17CloudflareREFINESAPR 6MetaREFINESMAR 2StripeREFINESFEB 12Gloaguen et al.REFINESJAN 28Lulla et al.SUPPORTSJAN 6Dae-Kyoo KimREFINESDEC 9SpotifySUPPORTSDEC 2BunREFINESNOV 24SpotifySUPPORTSNOV 12Shafin et al.REFINESNOV 6SpotifySUPPORTSNOV 2Yin et al.REFINESOCT 21Google CloudSUPPORTSSEP 23Google Cloud / DORAREFINESSEP 21Deng et al.REFINESJUL 10METRREFINESMAY 16OpenAIREFINESMAY 15CursorSUPPORTSMAR 13AirbnbSUPPORTS stage mapped adjacent stage signal Selection condition mapped→ explicit transition in canonical evidence mappingSUPPORTS / REFINES model implication
Agent runtime exceeds human labor before human steering disappears● Selection■ Execution■ Verification■ Coordination■ EconomicsREFINESCooperation has to earn its coordination cost● Selection→● Cooperation○ Specialization■ Context■ Verification■ Coordination■ Economics■ LearningREFINESEvals make selection pressure executable● Selection■ Verification■ Observability■ Economics■ LearningREFINESContext becomes an engineered workflow● Selection■ Context■ Execution■ Verification■ LearningSUPPORTSThe software factory selects for complete managed capabilities● Selection→● Cooperation○ Specialization■ Context■ Execution■ Verification■ Coordination■ Observability■ EconomicsREFINESThe Internet becomes part of the agent fitness function● Selection→● Cooperation■ Context■ Execution■ Coordination■ Observability■ EconomicsREFINESSpecialized coordination turns swarm activity into useful progress● Selection● Cooperation● Specialization■ Context■ Execution■ Verification■ Coordination■ EconomicsREFINESHuman expertise stays load-bearing while the agent executes● Apparition● Selection■ Context■ Execution■ Verification■ ObservabilityREFINESStripe extends coding-agent operation into infrastructure provisioning with provider integration, persistent context, and explicit safety boundaries● Selection→● Cooperation■ Context■ Execution■ Coordination■ Observability■ EconomicsREFINESCoding stops being the constraint● Selection→● Cooperation■ Context■ Execution■ Verification■ Coordination■ Observability■ EconomicsSUPPORTSLong-running agents require a runtime● Selection■ Execution■ Observability■ LearningSUPPORTSHuman attention becomes the concurrency ceiling● Selection→● Cooperation■ Execution■ Verification■ Coordination■ ObservabilityREFINESThe interface reorganizes around many agents● Selection→● Cooperation■ Context■ Execution■ Verification■ CoordinationSUPPORTSSpotify connects repository discoverability and platform context to autonomous migration outcomes● Selection→● Cooperation■ Context■ Execution■ Coordination■ EconomicsSUPPORTSAgent fitness becomes measurable at the environment boundary● Selection■ Context■ Execution■ Observability■ Economics■ LearningREFINESMeta reduced coding-agent exploration by pre-computing concise, quality-gated repository context with specialized agents● Selection● Cooperation→● Specialization■ Context■ Verification■ Coordination■ Observability■ Economics■ LearningREFINESStripe benchmarked coding agents on production-realistic full-stack integration work and exposed both long-horizon capability and recovery failures● Apparition● Selection■ Context■ Execution■ Verification■ Observability■ LearningREFINESContext files raise cost without significant task-success gains● Selection■ Context■ Verification■ EconomicsREFINESRepository instructions reduce agent runtime and output tokens● Selection■ Context■ EconomicsSUPPORTSTask-, phase-, and process-level agent specialization showed different strengths but no universally fitter configuration● Selection● Specialization■ Context■ Execution■ Verification■ Coordination■ LearningREFINESSpotify measures how verification feedback changes background-agent outcomes● Selection→● Cooperation■ Execution■ Verification■ ObservabilitySUPPORTSAgent infrastructure becomes strategically load-bearing● Selection■ Execution■ VerificationREFINESSpotify shows context engineering as a selection condition for background coding agents● Selection■ Context■ Execution■ VerificationSUPPORTSRole-specialized cooperation changes engineering fitness rather than uniformly improving it● Cooperation■ Context■ Verification■ CoordinationREFINESSpotify puts background coding agents into fleet-scale software maintenance● Apparition● Selection→● Cooperation■ Execution■ Verification■ Coordination■ Observability■ EconomicsSUPPORTSSeven agent frameworks showed distinct effectiveness, coordination, and cost trade-offs across three software-engineering task families● Apparition● Selection● Cooperation■ Execution■ Verification■ Coordination■ Economics■ LearningREFINESGoogle paired nested coding-agent loops with build, test, monitoring, and fleet automation for an ISA migration spanning 100K+ applications● Apparition● Selection■ Context■ Execution■ Verification■ Coordination■ Observability■ EconomicsSUPPORTSDORA finds AI outcomes depend on the surrounding engineering system, not AI adoption alone● Selection■ Context■ Execution■ VerificationREFINESLong-horizon repository work exposes both viable agentic action and a large reliability gap● Apparition● Selection■ Context■ Execution■ VerificationREFINESAI availability slowed experienced maintainers by 19% on familiar real-world tasks● Selection■ Context■ Verification■ EconomicsREFINESCodex launched as a cloud software-engineering agent that could independently edit repositories, execute checks, and return reviewable evidence● Apparition● Selection■ Context■ Execution■ Verification■ ObservabilityREFINESCursor introduced remote background coding agents that could run many delegated engineering tasks in parallel● Apparition■ Execution■ Coordination■ ObservabilitySUPPORTSValidation, context, and retries made an LLM migration viable at 3,500-file scale● Selection■ Context■ Execution■ Verification■ Observability■ Economics■ LearningSUPPORTS
● stage○ adjacent stage→ transition■ selection condition