AES / FIELD NOTES
Technical essays and operating principles for companies where people and AI agents work as one governed system.
Latest research
GitHub’s new npm, Dependabot and Actions controls reveal why agent runtimes need provenance, quarantine states and governed activation for every external capability.
Read article
GitHub’s latest Copilot controls show why enterprise agents need a shared policy plane, durable authority boundaries and evidence across every execution surface.
Read article
Why autonomous organizations need more than a vector database—and how to design memory agents can trust, update and act on.
Read article
MCP standardizes access to capabilities. An enterprise runtime must still own identity, policy, orchestration, evidence and accountability.
Read articleAES / 2026
Anthropic’s ART announcement is not evidence of autonomous science. It is evidence that AI-assisted discovery becomes more consequential when model work, scientific judgment and experimental validation operate as one research system.
Read article
The CMA’s draft could let qualifying AI assistants compete for placement on Google search-choice screens. The proposal is not final, but it makes web access, attribution and platform integration conditions of search distribution.
Read article
Microsoft’s disruption of EvilTokens changes the fraud assumption: control of a mailbox is no longer credible proof for a payment change. The service paired device-code phishing and token theft with AI-assisted mailbox analysis to identify trusted relationships and payment authority.
Read article
Google’s new GKE results show a three-region inference deployment approaching local-call throughput in one specific test. The enterprise question is whether a distributed allocation can become one serving fleet.
Read article
AgentCore Runtime V2 can reclaim released or cold memory during a session rather than billing against its peak allocation until termination. The economic result is workload-specific: V2 also has higher CPU and memory rates.
Read article
In a deep service graph, each retry claims scarce recovery capacity. The system needs one designated recovery point, not a fresh retry loop at every hop.
Read article
KDDI reports cutting RAG evaluation work by 75% by sampling a grid of file formats and media composition. The reusable lesson is to allocate testing effort across the input space that actually changes system behavior.
Read article
Sponsored Agents make a paid placement an optional, labeled conversation with a business. The immediate marketing question is whether product knowledge and attribution are ready for that surface.
Read article
MLCommons says MLPerf Endpoints will replace datacenter MLPerf Inference. The practical change is a benchmark that compares a deployed endpoint across the load and latency trade-offs buyers actually operate.
Read article
Lambda’s five-rack validation of NVIDIA DSX MaxLPS raised pure-inference throughput by 24% within a fixed 129-kW power budget. The result is not a general benchmark. It is operating evidence that dynamic power allocation belongs in AI capacity planning.
Read article
Microsoft AI’s draft code says future MAI models may give up generality, autonomy or capability to preserve human control. For enterprise buyers, the practical change is a new procurement question: which limits are built into the model, and do they apply to the model being bought?
Read article
Why faster delivery can generate more work, invisible rescue can preserve broken dependencies, and the next useful intervention may be a boundary rather than another instruction. An authorial working hypothesis by Maxim Perfiljev.
Read article
OpenAI reports that two engineers, assisted by Codex and GPT-5.5, moved most production traffic for Habitat from Python to Rust. The result does not make deferred rewrites universally safe. It shows how stable interfaces, production measurement and a controlled migration can change when a rewrite is worth doing.
Read article
AI can remove specialist handoffs and accelerate delivery. It can also remove the situated work through which a company produces its next reviewers, troubleshooters and decision-makers.
Read article
DiDi’s reported rise from 38% to 86% offers a narrow but powerful design lesson: a model cannot fairly verify an existing label while being invited to replace it.
Read article
Microsoft, the AFT and the UFT have created an optional contract standard that lets U.S. school districts turn selected AI safeguards into enforceable vendor obligations.
Read article
AlphaGenome Atlas precomputes predicted molecular effects for more than 9 billion single-letter DNA substitutions. The change is a new research data layer: it lets teams prioritize whole genomes before allocating scarce experimental validation.
Read article
Amazon and Qualcomm have announced a multi-generation AI infrastructure collaboration backed by initial purchase commitments and a warrant whose further vesting depends on orders and purchases. The structure makes compute, connectivity and supply-chain co-design part of hyperscaler competition.
Read article
Google and Cathay Pacific report an estimated 40% reduction in contrail warming impact across more than 80 trial flights that followed AI-informed routes. The result is promising operational evidence, not proof of fleet-wide performance.
Read article
OpenAI reports 3.1 eight-hour AI-runtime workdays for each human workday in its research organization. The figure is not an FTE or productivity measure, but it makes inference capacity a visible input to research operations.
Read article
Falling model costs do not automatically lower the cost of completed work. In many AI workflows, scarce capacity to test, reconcile, review and correct outputs becomes the binding production constraint.
Read article
A build-in-public account of why AES reduced its editorial supervision from seven daily checkpoints to two, preserved failures before recovery, and separated safety gates from constant intervention.
Read article
OpenAI's new model expands the scope and duration of computer-based work. The enterprise response should be trajectory governance, not another benchmark comparison.
Read article
JFrog’s AgentSecOps announcement treats the prompts, skills, plugins and MCP servers used by coding agents as supply-chain inputs. The important architectural shift is reproducible agent context—not a claim that versioning solves agent governance.
Read article
Anthropic’s Fable 5.1 and Mythos 5.1 share an underlying model, but their safeguards and access conditions require autonomous organizations to govern them as separate runtime capabilities.
Read article
Enterprise Frontier Safeguards proposes a cross-cloud division of responsibility: Anthropic supplies automated safety detection while customers retain telemetry, keys, logs and human review.
Read article
Broadcom has made AgentMinder generally available: a runtime authorization gateway for enterprise AI agents. The important move is not agent identity alone, but a fresh policy decision before each tool call.
Read article
Why autonomous organizations need a governed replay gate that tests proposed configurations against sealed historical executions before promotion.
Read article
Amazon Bedrock AgentCore Memory now evaluates identity claims, policy, operation and namespace before gateway requests reach persistent memory. The change makes memory access a runtime authorization decision—but only on configured gateway routes and supported operations.
Read article
Autonomous organizations need evidence generated at policy, effect and state boundaries—not an agent’s account of what it did.
Read article
A recovery architecture for autonomous organizations must restore evidence and unfinished work without automatically reviving revoked authority, stale observations or already completed external effects.
Read article
Why autonomous organizations need a canonical obligation ledger for work that survives the agent, workflow and policy decision that created it.
Read article
Nintex’s new Solutions container groups agents and deterministic automation around a business outcome. The important move is a shared lifecycle—not an agent feature.
Read article
Credentials should be execution artifacts, issued to an attested workload for one governed effect—not durable knowledge available to an agent.
Read article
Salesforce’s Headless 360 expansion exposes governed business capabilities to agents and applications. The important architectural change is not MCP itself, but keeping enterprise controls in the execution path.
Read article
Agent-generated code, policies, workflows and configurations should be promoted as verified artifacts—not activated directly from an execution.
Read article
Google Cloud’s financial-services preview treats research methodology as a reusable agent asset. The harder task is governing that asset as organizational policy.
Read article
A governed agent runtime needs a stack of identities: organizational role, verified workload, runtime instance, execution and authority grant.
Read article
Thomson Reuters has put its proprietary Thomson model into a multimodel CoCounsel architecture. The material control is not model ownership alone, but an auditable decision about which model may perform each task.
Read article
Per-action authorization cannot see the aggregate exposure created by concurrent autonomous work. Governed runtimes need a risk-budget ledger above the policy engine.
Read article
The announced partnership connects live court records to legal AI workflows. The operating question is how an organization turns each new filing into governed work without mistaking a draft for legal action.
Read article
Authentication identifies the requester. Authorization defines its scope. For consequential agent actions, neither establishes that the approved runtime was present at the moment of execution.
Read article
Emergency authority should be a bounded recovery state with independent dependencies—not a permanent administrator account waiting for a crisis.
Read article
Generative agents can remain probabilistic—provided that the organization’s routing, authority, recovery and commitment logic is reproducible.
Read article
Anthropic has made browser use, computer use, Skills and Files generally available on the Claude API. The significant enterprise boundary is not the model endpoint, but the customer-hosted browser executor.
Read article
A negative-decision ledger turns authorized silence, deferral, inapplicability and unresolved failure into accountable runtime evidence.
Read article
Pega Customer Engagement Studio turns conversational campaign briefs into validated, approvable executable rules. The important boundary is not the chat interface; it is the governed translation from intent to action.
Read article
OpenAI’s new cyber-capability safeguards show why monitoring for powerful agents must govern execution, consume capacity, and be able to stop work—not merely observe it afterward.
Read article
Identity and permissions say who may act and what they may do. Governed agent systems also need to decide why a particular action is allowed now.
Read article
Autonomous organizations need executable metapolicy to resolve conflicts among independently authored rules before an agent can act.
Read article
A2A’s move into AAIF aligns governance around agent and tool protocols. It does not create the execution continuity enterprises need across agent delegation and tool use.
Read article
Anthropic will watermark text from future Claude models globally. The signal can support disclosure, but autonomous organizations still need their own evidence of execution, approval and distribution.
Read article
A consequential agent decision should be governed as a logical read transaction: choose an isolation class, preserve its read manifest, validate it before effect, and rerun reasoning when the state has changed.
Read article
A renewable work lease gives an agent temporary standing to advance a task—and gives the organization a disciplined way to stop, replace and govern that agent when certainty is lost.
Read article
Autonomous organizations need a deterministic boundary between agent negotiation and commitments that can bind, change or end an organizational obligation.
Read article
Autonomous organizations need a work-admission protocol between an agent’s intent and its execution: reserve the operating envelope first, then decide whether work may begin.
Read article
OpenAI’s limited-preview Ultrafast API tier changes the enterprise question from how fast a model can respond to which agent steps are permitted to use speed—and under what authority.
Read article
Autonomous organizations need more than citations. They need a versioned dependency graph that can identify which beliefs, plans, approvals and effects are affected when evidence changes.
Read article
OpenAI has made Daybreak Blue and Red available through Amazon Bedrock for approved customers. The significant change is a governed access model: powerful cyber capability is conditional on organization, identity, purpose, controls and continuing review.
Read article
Deletion is not a database operation when knowledge has become embeddings, summaries, caches, agent memory and model inputs. Forgetting needs to be a governed runtime protocol.
Read article
OpenAI will disable individual-user sync connections in Enterprise and Edu from August 14. The change does not end administrator-managed sync; it makes ownership, scope and lifecycle of organizational knowledge an operational concern.
Read article
Long-running autonomous work needs more than deployable components. It needs an explicit compatibility envelope that binds workflow, agents, tools, data and policy to one business execution.
Read article
Meta’s local, open-weight agent model changes where capable agents can run. It does not provide governance; it makes portable governance necessary.
Read article
GitHub’s Copilot usage metrics API can now report recognized third-party agent activity by agent. It improves portfolio measurement, but usage counters are not operational accountability.
Read article
Autonomous organizations need civil time, elapsed time and authoritative order—and must govern confidence in each.
Read article
A governed autonomous organization should decide what can run in parallel from the business invariants each action can affect—not from the agents or workflows that happen to perform it.
Read article
Multiple agent approvals are meaningful only when the approving agents are independent across the failure domains that matter to a business effect.
Read article
Exactly-once is not a property of an agent run or a message. It is a governed claim about one approved business mutation.
Read article
A new Azure production study shows that agent workflows stress CPUs, GPUs and orchestration differently from conventional services. The operational consequence is clear: runtime placement and resource isolation must become governed system state.
Read article
A governed runtime should treat subagent permissions as derived authority: constrained at issuance, enforced at use, revocable through the grant graph, and accountable to both principal and actor.
Read article
Durable workflows preserve execution. Reconciliation loops determine whether an organization still needs work—and how to converge without turning drift into unsafe automation.
Read article
A new preprint shows why physical agents should translate procedures into inspectable, schema-constrained actions before execution.
Read article
When an autonomous workflow has already acted outside the system, recovery is not rollback. It is governed coordination toward a valid business outcome.
Read article
Tool authorization decides whether an agent can act. Information-flow control decides what data may travel through that action.
Read article
Article 50 transparency obligations now apply. For autonomous organizations, the practical task is to evaluate disclosure against each execution context and retain evidence that the required control ran.
Read article
Long-running autonomous processes need explicit semantics for policy change. Treat the policy revision as execution state, and make every transition visible, approved and auditable.
Read article
For a governed autonomous organization, “stop” must mean a verifiable runtime state: no new work, bounded authority, known side effects and explicit unresolved exceptions.
Read article
Governed agent operations need separate records for observability, recovery and lineage. A trace can explain a run; it cannot safely resume one.
Read article
Cognizant’s new EMEA AI Unit is an organisational move, not a new platform. Its build-deploy-run model makes the enduring question explicit: who operates a blended human-agent workflow after launch?
Read article
An agent’s plan is not a durable right to act. For consequential work, authorization must be checked again at the boundary where an external effect commits.
Read article
Synopsys’ new evaluation workflows for debug and implementation closure show why autonomous engineering must govern the complete run—and retain evidence at every transition.
Read article
A credible agent marketplace must connect role, version, evaluation evidence, authority boundaries, safe trials and operational outcomes.
Read article
Why enterprise agents need durable identity, delegation chains, scoped capabilities, short-lived credentials and revocation outside the model session.
Read article
The difference between a company using AI tools and an organization designed for people and intelligent agents to work as one team.
Read article
Why the number of agents is the wrong metric—and what a coordination layer must provide for reliable enterprise execution.
Read article
How X10 creates a domain ecosystem for AI-native initiatives while AES remains the central product and operating focus.
Read articleBUILD WITH AES
AES connects strategy, tasks, organizational memory, knowledge, agents, people and approvals in one execution environment.
