AGENTHESIS — Agentic AI Researchhttps://agenthesis.techResearch into intelligence that acts.enWhen does an agent actually need another agent?https://agenthesis.tech/research/delegation-is-not-freehttps://agenthesis.tech/research/delegation-is-not-freeSun, 01 Nov 2026 12:00:00 GMTThe overlooked cost of delegation in tool-heavy workflows.CoordinationThe harness may matter as much as the model.https://agenthesis.tech/research/harness-model-pairinghttps://agenthesis.tech/research/harness-model-pairingSun, 01 Nov 2026 12:00:00 GMTWhy changing the agent loop can swing the result without changing model weights.EvaluationThe bug you introduce today may sabotage tomorrow.https://agenthesis.tech/research/featurearc-inherited-defectshttps://agenthesis.tech/research/featurearc-inherited-defectsSun, 01 Nov 2026 12:00:00 GMTLong-horizon coding benchmarks expose inherited defects.ReliabilityPassing tests are not necessarily good tests.https://agenthesis.tech/research/tests-that-detect-bugshttps://agenthesis.tech/research/tests-that-detect-bugsSun, 01 Nov 2026 12:00:00 GMTBehavioral adequacy is a different objective from test execution.VerificationKnowing a tool action is risky is not enough.https://agenthesis.tech/research/agentic-safety-outcomeshttps://agenthesis.tech/research/agentic-safety-outcomesSun, 01 Nov 2026 12:00:00 GMTSafety should be measured by actual system effects.SafetyWhat if agent teams could reorganize mid-task?https://agenthesis.tech/research/evolving-agent-teamshttps://agenthesis.tech/research/evolving-agent-teamsFri, 30 Oct 2026 12:00:00 GMTAdaptive roles can sometimes escape an optimization plateau.CoordinationWhat counts as evidence that an agent finished?https://agenthesis.tech/research/proof-of-completionhttps://agenthesis.tech/research/proof-of-completionFri, 30 Oct 2026 12:00:00 GMTObservability is not the same as a trustworthy completion claim.VerificationWhy code-generation agents need explicit contracts.https://agenthesis.tech/research/contracts-for-code-agentshttps://agenthesis.tech/research/contracts-for-code-agentsFri, 30 Oct 2026 12:00:00 GMTFile ownership and compile checks can improve buildability.EngineeringAgents can form informal rules around scarce compute.https://agenthesis.tech/research/scarce-compute-social-ruleshttps://agenthesis.tech/research/scarce-compute-social-rulesFri, 30 Oct 2026 12:00:00 GMTResource allocation becomes a coordination problem with surprising behaviors.CoordinationAgents cooperate better when they can see each other.https://agenthesis.tech/research/shared-git-peer-awarenesshttps://agenthesis.tech/research/shared-git-peer-awarenessWed, 28 Oct 2026 12:00:00 GMTAn observable shared workspace may outperform silent parallelism.Coordination