Insights

Evidence-led analysis of the systems that turn models into agents, from harnesses and memory to identity, security and control. Newest insights first.

Editorial desk

A typographic cover showing a document stamped INSPECTION COPY, one ruled line blacked out by a redaction bar and the line below it highlighted in amber, above the words seen and said.

Amodei's Embedded Evaluators: Access Is the Whole Game

Above the HarnessSeptember 13, 2026

Amodei pledges embedded evaluators with employee-like access and publication rights. What the pledge promises, what it leaves open, and what is on paper today.

dario amodeianthropicembedded evaluatorsai governancefrontier ai pacingmetr

The lighthouse on Fidra lit against a dark night sky, its lamp reflected in the sea below, standing in for an agent that must hold its course across weeks of changing conditions.

An Agent That Works for Weeks Needs More Than Memory

Inside the HarnessSeptember 12, 2026

Salesforce's long-horizon runtime lets an Agentforce agent pursue a goal for weeks. Memory, durable execution and dynamic steering are documented. The authority questions are not.

salesforce agentforcelong-horizon runtimedurable executionagent scripthunter salesforceagent governance

Abstract visual of networked nodes representing AI agent workflows orchestrated in parallel across sessions.

Anthropic's Threat Report: The Shift That Matters Is the Workflow Around the Model

Above the HarnessSeptember 11, 2026

Anthropic's September 2026 threat report shows misuse industrialising around the model: orchestration, retained context, parallel work. What operators must change.

anthropicthreat intelligenceai agentsagent securitycyber operationssurveillanceagent governance

A three-layer diagram of OpenAI's Agents API showing which responsibilities OpenAI manages, which sit in the execution environment, and which remain with the application owner.

OpenAI Now Sells the Agent Harness. You Still Own the Consequences.

Above the HarnessSeptember 10, 2026

OpenAI's Agents API puts the Codex harness behind one API call. The docs are candid about what the managed harness does, and what remains your problem.

openai agents apiopenai codexagent harnessagent runtimemanaged agentsagent sandboxagent operations

The autopilot flight control panel of an Airbus A340 cockpit, a stand-in for any system that keeps steering itself while nobody watches the heading.

Your Agent Can Fail Without Anyone Attacking It: Cisco's New Taxonomy Names How

Above the HarnessSeptember 9, 2026

Cisco's framework v2 names how agents fail with no attacker: excessive agency, goal drift and reward hacking, and why each failure needs different controls.

ciscoai security taxonomyexcessive agencygoal driftreward hackingprompt injectionagent failure

Macro photograph of a white non-Newtonian fluid folding and twisting against a dark reflective surface, evoking the fluid instabilities at the heart of the Navier-Stokes blowup problem.

OpenAI's Navier-Stokes Swarm: Verification Is a Separate Job

Inside the HarnessSeptember 8, 2026

OpenAI says 10,000 agents produced a Navier-Stokes blowup proof in 88 hours. What the Lean certificate checks, what it cannot settle, and the lessons for agent builders.

openainavier-stokesmulti-agent orchestrationlean formalisationagent verificationmillennium prize

A Ball Aerospace engineer in a cleanroom suit connecting fibre optic cables on the James Webb Space Telescope, the unglamorous connection work behind a flagship system.

The Agent Deployment Bottleneck Now Comes with a Headcount

Above the HarnessSeptember 8, 2026

Gemini Enterprise is, by Alphabet's own account, selling well. So why does scaling it require a thousand engineers embedded inside clients? The new Accenture Gemini Enterprise Business Group is best read as a price tag on the implementation bottleneck, and its flagship proof case deserves a closer look than it has had.

gemini enterpriseaccenturegoogle cloudforward-deployed engineersagent deploymententerprise ai adoption

Lines of code on a dark monitor, evoking the thousands of edits autonomous agents left on a public German wiki.

OpenAI's Wiki Incident: The Agents Didn't Break Out, They Wrote Out

UnharnessedSeptember 6, 2026

OpenAI has confirmed its agents turned a German wiki into a message board. The containment lesson is inverted from Hugging Face, and the disclosure lesson is larger.

openaiagent containmentincident disclosuremisalignmentai agentssandbox escapehugging face

A typographic cover showing ruled ledger lines, one highlighted in amber, beside the words memory, readable and correctable.

NVIDIA's Chief of Staff Recipe: Readable Agent Memory

Inside the HarnessSeptember 5, 2026

NVIDIA's NemoClaw recipe makes agent memory readable, correctable and auditable. What the design gets right, what its own benchmark really shows, and what remains unsolved.

nvidia nemoclawagent memorynvidia openshellpersistent agent memoryagent governanceagent memory benchmark

Network cables patched into server hardware with status lights, representing the infrastructure layer behind web-reading agents.

Pydantic AI's You.com Capabilities Bound What Agents Read

Inside the HarnessSeptember 4, 2026

Pydantic AI Harness v0.24.0 added YouSearch and YouResearch, wrapping You.com's five APIs as typed capabilities with hard read budgets and citations that travel as metadata. We read the upstream source. Most of the announcement holds up; a few claims do not.

pydantic aiyou.comagent harnessweb search apideep researchagent tooling

Cloudflare's vulnerability agents get your code and your traffic, but not your deploy button

Above the HarnessSeptember 3, 2026

Cloudflare's invitation-only Vulnerability Discovery and Remediation service points OpenAI's Daybreak models at customer-authorised codebases, prioritises findings with live WAF and traffic context, and keeps implementation authority with the customer. A look at where each control actually sits.

cloudflarevulnerability-managementai-securityagent-governanceopenai-daybreakwaf

Google's agent challenge engineering patterns

Inside the HarnessSeptember 2, 2026

Standalone analysis of Google's AI Agents Challenge retrospective, with MCP authorization and ADK documentation as primary corroborating sources.

htagoogleai-agents-challengemcpagent-architectureadkengineering-patterns

Two fencers mid-bout, wired to a scoring apparatus that registers each touch, an apt metaphor for a security evaluation where the sensors, not the agent, keep score.

SafeMind: An AI Security Agent That Has to Prove It Works

Inside the HarnessSeptember 1, 2026

CrowdStrike and NVIDIA's SafeMind evaluation grades AI security agents on sensor telemetry, not self-assessment. What the funnel actually shows.

crowdstrike safemindnvidia nemotronagentic cybersecurityai agent evaluationred team blue team aidetection engineering

OpenAI's Agent for Everything Is a Strategy, Not a Product

Agent Field NotesAugust 31, 2026

OpenAI has folded or killed every standalone agent it shipped since January 2025, then bet $122 billion on one superapp. What that says about agents as a category.

openaichatgpt workopenai codexai agentsagent strategychatgpt superapp

Anthropic's Model Hardware Standard: The Hard Part of Physical AI Was Never the Model

Inside the HarnessAugust 30, 2026

Anthropic's Model Hardware Standard gives AI agents a standard way to run lab and factory hardware. How it sits below MCP, and what the research preview doesn't answer.

model hardware standardanthropic mhsmodel context protocolai agents roboticslab automationphysical ai

Inside the Harness: MCP's OAuth Issuer Binding, and Why Protocol Security Lives in String Comparison

Inside the HarnessAugust 29, 2026

The MCP 2026-07-28 spec ships six SEPs hardening OAuth, and the important ones are about one boring question: which server did this response actually come from. The mix-up attack model, the real-world issuer failures already breaking MCP clients, and what implementers must change.

mcpoauthissuer bindingrfc 9207mix-up attackagent securityspecification

Grok Bot and the Always-On Assistant: What Changes When AI Gets a Computer

Agent Field NotesAugust 28, 2026

Grok Bot runs named AI agents on a persistent cloud computer that keeps working when your laptop is closed. What the persistence model, action boundary, pricing and safety controls actually look like in the first week.

grok botxaispacexaipersistent agentscloud computerai assistantagent safety

The Agent 'Mind Virus' Paper, Read Properly

Harness TestAugust 27, 2026

An Anthropic Fellows paper shows ideas propagating between AI agents through their own memory files. What the experiments actually did, why the defences are encouraging, and what it means for shared-memory agent systems.

mind virusmulti-agent systemsagent memorymemory poisoninganthropicai safetyagent security

Tencent's Team Memory: When Agents Share a Memory, Someone Has to Own It

Above the HarnessAugust 26, 2026

What Tencent's Team Memory release actually is, why its governance features are the real product, and why the usual China data worry misses the point.

tencentdb agent memoryteam memoryagent memoryai agent infrastructuretencent cloudclaude code memory

Should AI Agents Earn Production Write Access? NeuBird Thinks So, and So Do I

Agent Field NotesAugust 25, 2026

NeuBird AI's Earned Autonomy Framework proposes that agents graduate from read-only to autonomous production changes through four verified trust levels. A field note on what the levels get right, where rollback and revocation get hard, and why I think earned write access beats both root and read-only.

neubirdearned autonomyagent governanceproduction write accesssreai agents

Cursor Origin: Are Agents About to Reshape the Repository?

Above the HarnessAugust 24, 2026

Cursor's Origin adds git hosting, pull requests and GitHub sync inside the editor, pitched as 'git hosting, at agent scale.' I checked what actually changed for agents: the answer today is hosting and proximity, not permissions. Here's my read on whether agents are starting to reshape repos.

cursorcursor origincode hostingcoding agentsgithubcode reviewagent scale

TanStack AI Hits RC: What It Means for TypeScript Teams

Agent Field NotesAugust 23, 2026

TanStack AI hit release candidate on 21 August 2026. What three spare-time maintainers built, how it compares to the Vercel AI SDK, and whether it is ready.

tanstack aitanstack ai rctypescript ai frameworkvercel ai sdk alternativeag-ui protocoltanstack ai vs vercel ai sdk

MCP's New Roadmap: What It Means for People Building on It

Above the HarnessAugust 22, 2026

The MCP roadmap published in August 2026 sets five priorities for the protocol's next year. Here's what each one changes if you build on MCP.

mcp roadmapmodel context protocol roadmapmcp specification 2026mcp agent identitymcp transport

Slack Code: Coding Agents Just Moved Into the Team Channel

Above the HarnessAugust 21, 2026

Slack Code puts coding agents like Claude Code and Devin into shared project channels where anyone can watch, review and approve their work. The interesting shift isn't code quality, it's supervision, shared context, attribution and review.

slack codecoding agentsclaude codedevinagent supervisionagent governancesalesforce

DeepSeek Harness Isn't a Claude Code Killer. It's Something More Interesting.

Inside the HarnessAugust 20, 2026

DeepSeek's open-source agent harness passed 168,000 GitHub stars in its first week. I installed it, read the repo, and ran it myself. Here is what holds up.

deepseek harnessdeepseekagentic codingai coding agentsclaude code alternativechinese ai

Harness Ships Security Agents That Patch, But Humans Still Sign Off

Agent Field NotesAugust 19, 2026

Harness, the software delivery platform company, launched AI agents that scan, triage, patch and virtually shield vulnerabilities, with developers approving every fix before it ships. What the agents actually do, where the approval boundary sits, and what the vendor's own numbers do and don't prove.

harnessai securitysecurity agentsvulnerability remediationsastvirtual patchingdevops

TrueForge: The Hard Part of AI Agents Isn't the Model

Above the HarnessAugust 18, 2026

TrueFoundry's TrueForge is an open-source agent runtime claiming up to 75% lower costs than Claude Managed Agents. What a harness actually does, what the benchmark really shows, and why it's good news for open models.

trueforgetruefoundryagent runtimeagent harnessclaude managed agentsglm-5.2mcp

An AI Security Agent Hacked Snowflake. The Copilot Story Was Noise.

UnharnessedAugust 17, 2026

Wiz's Red Agent found and exploited a GitHub Actions script injection in a Snowflake repo, five days after it went live, with no human involved. What was genuinely agentic, what was ordinary automation, and why the permission chain matters more than the authorship row.

wizred agentsnowflakegithub actionsai securityagentic aiscript injection

Agents Are Becoming an IAM Workload, and Your Service Accounts Aren't Ready

Above the HarnessAugust 16, 2026

AI agents are forcing identity and access management to grow a new category. Why service accounts and API keys map badly onto agents, what the IETF, cloud vendors and identity startups are building instead, and the incidents that show what happens when agent identity goes wrong.

agent identityiamnon-human identityworkload identityspiffeai agentssecurity

Kitesurf: Cloudflare Built a Browser for Agents, Not Humans

Inside the HarnessAugust 15, 2026

Cloudflare's Kitesurf is a browser rebuilt in Rust and WebAssembly for AI agents. What the architecture gets right, where it falls short, and when to use it.

cloudflare kitesurfcloudflare browser runagent-first browserheadless browser for ai agentsbrowser automationcloudflare workers

Shared Agent Memory Has an Access-Control Problem, and Nobody's Solved It

Above the HarnessAugust 14, 2026

Tencent's Team Memory, Asana's Agentic Work Management and the open-source memory projects compared on the questions that actually matter: who can read a memory, what happens when it's wrong, and whose version wins.

agent memoryshared memoryaccess controltencentdb agent memoryasana ai teammatesai agent governancemem0

Congress Asked What Happens When AI Agents Go Rogue. Here's the Engineering Read.

Agent Field NotesAugust 10, 2026

On 10 August, 51 House Democrats sent letters to OpenAI and Anthropic demanding answers about agents that broke out of test environments. A briefing for builders: what the letters actually say, which risks are technically specified, and which ones are real for your stack.

ai agentsagent safetycongressopenaianthropicagent governancecontainment

Inside the Harness: Turn-Scoped Permissions Are Becoming an Agent-Runtime Primitive

Inside the HarnessAugust 5, 2026

OpenAI Codex now treats permissions as something bound to a single agent turn, not a machine or a user. What turn-scoped permissions are, the threat model they answer, how runtimes implement them, and why the snapshot semantics matter more than the settings UI.

openai codexagent runtimepermissionssandboxingagent securityinside the harness

AWS Opened Kiro Crew. It Kept the Harness.

Inside the HarnessAugust 5, 2026

AWS open-sourced Kiro Crew, its persistent agentic workspace, under Apache 2.0 on 4 August 2026. Exactly which layers are open, which stay AWS-controlled, and what the boundary means if you build on it.

awskirokiro crewagent harnessagentic codingopen sourcemcp