The forgotten operating system
What's unlocking the latest agent capabilities? Kanban boards. Explicit handoff protocols. Delegation with scope and boundaries. Root-cause analysis. Memory and learning loops. Specs and skill files as durable operating protocols.
None of this is new. Frontier firms are rediscovering and applying late-20th-century management practices that many organisations let atrophy.
Of course, lean and agile weren't invented to manage agents. They were invented to manage work across people and machines under uncertainty, where production was faster than coordination could keep up; the very conditions agent-infused work now operates in. Visible work-in-progress on a board. Root-cause analysis when something fails. Memory that survives the task that generated it. With an operational substrate, agent throughput compounds into capability that can connect across the horizontal work of an organisation.
Will "dusting off" lean and agile practices fix the absorption gap caused by the exponential capability of agents? Here's where the evidence is stacking up.
-
Liao et al. (Meta)·evidence
Storing an agent's exploration history as a queryable database, not disposable logs, reached target speedups 10x faster.
The paper treats the artifacts, tool outputs, rewards, and causal lineage an agent generates over hundreds of steps as governed, queryable database state rather than disposable session logs. Grounded in a production accelerator-kernel optimiser at Meta, cross-session reuse of that stored experience reached a target speedup about 10x faster and at lower token cost. It is a quantified production instance of the Layer 3 claim that memory which survives the task compounds agent throughput into durable capability.
-
Noah Zweben, Anthropic·illustration
Anthropic's production agents act under their own scoped service accounts, so each action logs against the agent rather than a borrowed human credential.
Claude Tag issues each agent a service account in every connected system, scoped to a compartment such as a Slack channel or workspace, so the access question becomes what an agent may do within its boundary rather than what a user may do. The per-agent audit trail lets a failure be traced to the agent that caused it, which is the accountability bonding the framework's moral-crumple-zone design target asks for so blame does not fall on the human operator. It supports the Layer 3 claim that delegation with scope and boundaries is part of the operational substrate, not decoration.
-
Google Cloud·illustration
Google publishes the Open Knowledge Format, a vendor-neutral spec for knowledge as a directory of frontmatter-tagged markdown an agent can query and edit.
Google has published Open Knowledge Format v0.1, a vendor-neutral specification representing organisational knowledge as a directory of frontmatter-tagged markdown documents that an agent can query and edit, with optional index and log files and a one-page spec in the GoogleCloudPlatform/knowledge-catalog repository. It is the second hyperscaler to formalise markdown directories as the operating substrate for agent-readable knowledge. The move corroborates the site's reading that the medium of agentic coordination is markdown, and that specs and skill files are becoming durable operating protocols across vendors.
-
Anthropic·evidence
Anthropic's analytics agents answer ~95% of internal business queries at ~95% accuracy; without skills, accuracy never exceeded 21%.
The decisive figure is the eval delta: without skills the agents did not exceed 21% accuracy, rising consistently above 95% once skills were added.
-
Kieran Klaassen, Every·illustration
With a good plan and the right context, the build phase becomes routine and human effort shifts to the two ends of the loop, ideation and polish.
The original compound-engineering loop (brainstorm → work → review → compound → repeat) extends to eight steps with ideate and brainstorm prepended and polish appended.
-
OpenAI Developers Cookbook·illustration
Codex's Goals feature replaces ask-and-wait prompting with a persistent contract the agent works until a named verification surface confirms completion.
-
Gottweis et al., Nature·evidence
A Gemini-based multi-agent system produced hypotheses that outranked the curating experts' best guesses on the hardest goals, validated in three wet labs.
-
Brandon Gell & Willie Williams, Every·evidence
Every rolled out a personal agent per employee, abandoned it within weeks, rebuilt around shared coworkers with explicit roles, group access, and group memory.
-
John Cutler, The Beautiful Mess·evidence
TBM 422: Exception, Presence, Delegation as the three-motion working repertoire of management as infrastructure.
-
Mike Krieger, Anthropic CPO·evidence
"Anthropic is shipping our harness strategy, not a product."
-
Helm, Priebe, Duderstadt, arXiv cs.MA·evidence
Adaptive control charts extend lean/SPC to multi-agent LLMs, exposing a fundamental learn-versus-monitor tradeoff.
-
Li et al., arXiv cs.MA·evidence
Constraint drift formalised: safety-critical constraints degrade through memory, delegation, communication, audit, and optimisation.
-
Mario Hayashi·illustration
Developer applies Jidoka, Poka yoke, Andon, Muda, and Five Whys by name to AI agent failures.
-
David Rice, People Managing People·evidence
"AI is the reward for operational discipline, but it won't create it."
-
John Cutler, The Beautiful Mess·illustration
AI clears the blockers built from workflow friction and accelerates the ones built from dysfunction.
Cutler splits organisational blockers in two. Where a practice is held back by poor signal visibility, workflow friction, or missing scaffolding, AI removes the drag and can sustain something a team wanted but could never afford; where it is held back by low voice safety, misaligned incentives, or fear, the same tooling makes the avoidance and performance theatre cheaper to automate. This scope-limits Layer 3: rediscovering lean and agile practices absorbs agent capability only when the organisational substrate already favours improvement, and the same infrastructure accelerates dysfunction when it does not.
-
Rohit Krishnan, Strange Loop Cannon·evidence
Blended AI councils kept only a quarter of the good ideas a single model raised.
Krishnan decomposed model answers into atomic idea cards, clustered them, and had two blind judges rate which survived under different council structures. A peer-review round behaved as a consensus detector, promoting ideas several models shared while dropping about three quarters of the useful ideas only one model had raised, and the structure that recovered them stored and ranked each contribution before any synthesis. This is the fourth multi-agent result, after Assign-All, Too Many Specialists, and CEAD, showing that coordination structure must be designed and evaluated per problem, and it points the remedy at memory that survives the task rather than at more deliberation.
-
SkillSafetyBench authors, arXiv cs.AI / cs.MA·evidence
Even with benign user requests, task-relevant skill materials can steer agents toward unsafe actions.
Planting instructions in the skill materials an agent reads mid-task can consistently induce unsafe actions on otherwise benign requests, with failure patterns that vary by risk domain, attack method, and scaffold-model pairing. That is direct counter-evidence to the framework's assumption that markdown skill files are neutral, durable carriers: the properties that make them durable (text, addressable, transferable, consumed downstream) are what make them a transferable attack surface. Layer 3 has no clean late-20th-century antecedent for the skill-file trust discipline this implies; the nearest analogue is software supply-chain hygiene, itself still being worked out.
-
John deVadoss, arXiv cs.MA·evidence
CEAD: design quality dominates governance in agent architecture outcomes across a 10K-task benchmark.
Capability-aligned design beats control-heavy / design-poor architectures (70.6% vs 50.8% safe success); pulls the framework toward design first, governance second.
-
Masters & Albrecht et al., DeepFlow / arXiv·evidence
Assign-All baseline outperforms chain-of-thought management 0.502 to 0.313 across twenty enterprise workflows.