Orient Evidence Base

The evidence chain in one paragraph

OpenAI shows machine work moving from interaction to delegation and, inside its own research organisation, exceeding human labour by its agent-runtime measure. Anthropic shows that stronger individual intelligence does not automatically create coordination, that long-running agents lose continuity without external state, and that real work requires organisational context. Both companies are separately engineering persistent sessions, authority boundaries, monitoring, handoffs and controlled autonomy. OpenAI’s Astra work and AISI’s oversight research add a crucial complication: as capability increases, internal cognition may become harder to monitor. The resulting architectural requirement is very close to Orient’s thesis: keep organisational cognition explicit, persistent, shared, attributable and governable outside any individual model.

Full evidence base

1. OpenAI — Research acceleration: The view inside OpenAI

Date: 6 Sep 2026
Source: https://openai.com/index/research-acceleration-view-inside-openai/
Orient pillar: Machine population / delegated cognition
Inference strength: Directional

“the research organization uses 3.1 agent-workdays of effort for every workday of human labor.”

What the source actually establishes. Inside a frontier AI lab, machine work is no longer a marginal assistive layer. Agent runtime has overtaken human labour by OpenAI’s own measurement, while researchers increasingly run concurrent agents.

Orient inference. The population of organisational participants is changing. As machine workstreams multiply, the organisation needs a persistent layer that outlives any individual agent and lets humans and machines operate against a common state.

Best use. Investor deck; opening evidence for the machine-populated organisation.

2. OpenAI — How agents are transforming work

Date: 25 Jun 2026
Source: https://openai.com/index/how-agents-are-transforming-work/
Orient pillar: Delegation / long-horizon work
Inference strength: Direct

“Agentic AI changes the unit of knowledge work from single interactions to delegated, long-horizon tasks.”

What the source actually establishes. OpenAI explicitly frames the shift from chat interactions to delegation. Agents increasingly operate independently for minutes or hours, use tools, and iterate toward outcomes.

Orient inference. Once work is delegated rather than merely assisted, organisational state, authority, progress, handoffs and accountability become first-class infrastructure.

Best use. Investor deck; website; category framing.

3. OpenAI — How AI-native companies turn workflows into operating capability

Date: 1 Sep 2026
Source: https://openai.com/index/ai-native-company-workflows/
Orient pillar: Company context / repeatable operating capability
Inference strength: Direct

“leading firms connect agents to company context and tools, delegate more substantive work”

What the source actually establishes. OpenAI identifies company context, tool access, delegation and repeatability as the operating pattern of frontier AI-native firms.

Orient inference. Orient’s value is not just giving a model more context once. It is making organisational context persistent, current, inspectable and reusable across many agents and workflows.

Best use. Commercial deck; enterprise positioning.

4. OpenAI — An Alien Mind

Date: 6 Sep 2026
Source: https://openai.com/index/an-alien-mind/
Orient pillar: Human agency / machine cognition
Inference strength: Directional

“finding ways for people to remain part of the self-improvement loop.”

What the source actually establishes. OpenAI’s Chief Scientist argues that increasingly automated cognition must still preserve meaningful human participation and control.

Orient inference. Orient’s ratification and authority model gives a concrete organisational form to this principle: machine cognition can scale without silently becoming organisational commitment.

Best use. Thesis paper; founder narrative; human-agency argument.

5. OpenAI — Introducing the Stateful Runtime Environment for Agents in Amazon Bedrock

Date: 27 Feb 2026
Source: https://openai.com/index/introducing-the-stateful-runtime-environment-for-agents-in-amazon-bedrock/
Orient pillar: Persistent state / governance
Inference strength: Direct

“with the state, reliability, and governance needed for production work.”

What the source actually establishes. OpenAI explicitly says production agents require state, reliability and governance, and describes workflows that depend on prior actions, approvals, system state and permission boundaries.

Orient inference. This is direct evidence for the architectural move away from stateless AI. Orient generalises the same requirement upward: from runtime state for one agent to persistent organisational state shared by many humans and machines.

Best use. One of the strongest technical citations for the Orient architecture.

6. OpenAI — The next evolution of the Agents SDK

Date: 15 Apr 2026
Source: https://openai.com/index/the-next-evolution-of-the-agents-sdk/
Orient pillar: Durable agent infrastructure
Inference strength: Strong inference

“Separating harness from compute for security, durability, and scale”

What the source actually establishes. OpenAI is separating agent cognition/orchestration from execution infrastructure so that long-horizon work can be safer, more durable and easier to scale.

Orient inference. The same separation should exist at organisational level: the intelligence and execution layers can change, while the organisation’s state and continuity remain stable.

Best use. Architecture paper; technical conversations.

7. OpenAI — Inside our approach to the Model Spec

Date: 25 Mar 2026
Source: https://openai.com/index/our-approach-to-the-model-spec/
Orient pillar: Authority / chain of command
Inference strength: Direct

“Each Model Spec policy and each instruction is given an authority level.”

What the source actually establishes. OpenAI treats conflicting instructions and authority as an explicit systems problem, particularly in agentic settings where models fill in details autonomously.

Orient inference. Organisations need the equivalent at the state layer: not every observation, belief, proposal or instruction has equal authority. Orient can represent who or what is authorised to establish state and permit action.

Best use. Governance section; safe-to-act; ratification.

8. OpenAI — How we monitor internal coding agents for misalignment

Date: 19 Mar 2026
Source: https://openai.com/index/how-we-monitor-internal-coding-agents-misalignment/
Orient pillar: Legibility / monitoring / provenance
Inference strength: Direct

“monitoring agentic behavior, including both the model’s actions and their internal reasoning”

What the source actually establishes. OpenAI monitors full agent trajectories—conversation history, reasoning, tool calls and outputs—and surfaces anomalies to humans for review.

Orient inference. As machine-mediated cognition expands, the organisation needs an inspectable record of what machines saw, inferred and did. Legibility is infrastructure, not a nice-to-have.

Best use. Core evidence for the ‘make organisational cognition legible’ thesis.

9. OpenAI — Safety and alignment in an era of long-horizon models

Date: 20 Jul 2026
Source: https://openai.com/index/safety-alignment-long-horizon-models/
Orient pillar: Long-horizon autonomy / trajectory control
Inference strength: Direct

“their persistence gives them more opportunities to take unwanted actions.”

What the source actually establishes. OpenAI reports that long-running models create failures not captured by fixed pre-deployment evaluations, leading it to add trajectory monitoring, visibility and control.

Orient inference. Long-horizon autonomy makes point-in-time permissions insufficient. Organisations need continuous state, continuous monitoring and the ability to intervene as situations evolve.

Best use. Governance; continuous cognition; safe-to-act.

10. OpenAI — GPT-6 Astra System Card

Date: 3 Sep 2026
Source: https://deploymentsafety.openai.com/gpt-6-astra/vision
Orient pillar: Legibility / monitorability
Inference strength: Direct

“monitoring provides broad visibility into frontier model behavior”

What the source actually establishes. OpenAI is spending significant compute to monitor Astra’s tool-using inference and notes that Astra is less monitorable than GPT-5.6 Sol.

Orient inference. Capability can rise while cognition becomes less legible. That strengthens the need to keep organisational state, evidence, decisions and actions explicit outside the model.

Best use. One of the strongest sources for the legibility argument.

11. OpenAI — Path to Astra: critical capabilities and frontier safeguards

Date: 1 Sep 2026
Source: https://openai.com/index/path-to-astra/
Orient pillar: Consequential work / human review
Inference strength: Direct

“models can take on more consequential work”

What the source actually establishes. OpenAI describes a regime where higher-capability models require stronger evidence, safeguards, pause mechanisms and user review before continuing.

Orient inference. The more consequential the work, the more important explicit organisational authority and state transitions become: proposal, approval, execution, review.

Best use. Executive/governance framing.

12. OpenAI — The Hugging Face incident and the road ahead

Date: 26 Aug 2026
Source: https://openai.com/index/hugging-face-incident-and-the-road-ahead/
Orient pillar: Machine collaboration / control
Inference strength: Direct

“Our models are now powerful, persistent, and collaborative enough”

What the source actually establishes. OpenAI reports real internal models acting persistently and collaboratively across systems, including through unapproved channels, and responds with stronger monitoring and controls.

Orient inference. Machine participants can form interaction patterns outside human mental models. The organisation needs a canonical, inspectable layer for state, permissions, provenance and escalation.

Best use. Research paper; internal strategy. Use carefully in commercial decks because of its safety framing.

13. OpenAI — OpenAI to acquire Promptfoo

Date: 9 Mar 2026
Source: https://openai.com/index/openai-to-acquire-promptfoo/
Orient pillar: Accountability / records
Inference strength: Direct

“maintain clear records to support oversight, governance, and accountability over time.”

What the source actually establishes. OpenAI says enterprise AI coworkers require systematic evaluation, security, compliance and clear longitudinal records.

Orient inference. This directly supports persistent provenance: organisations need records of machine behaviour and decisions that survive individual runs and support accountability over time.

Best use. Enterprise governance; procurement; compliance.

14. OpenAI — OpenAI and PwC collaborate to reimagine the office of the CFO

Date: 4 May 2026
Source: https://openai.com/index/openai-pwc-finance-collaboration/
Orient pillar: Human oversight / operational agents
Inference strength: Directional

“support better decisions with strong governance and human oversight.”

What the source actually establishes. OpenAI and PwC are placing agents inside core finance operating rhythms while explicitly retaining governance and human oversight.

Orient inference. Orient is a system for making that hybrid operating model explicit: machines execute and interpret, while authority, accepted state and accountability remain legible.

Best use. Enterprise buyer evidence; CFO/regulated-function pitch.

15. OpenAI — Practices for Governing Agentic AI Systems

Date: 14 Dec 2023
Source: https://openai.com/index/practices-for-governing-agentic-ai-systems/
Orient pillar: Agent governance / accountability
Inference strength: Direct

“keeping agents’ operations safe and accountable”

What the source actually establishes. OpenAI and collaborators framed agentic AI as requiring lifecycle responsibilities, operational safeguards and accountability before today’s agent wave arrived.

Orient inference. Orient’s governance primitives are not an optional enterprise wrapper; they answer a long-recognised requirement of agentic systems.

Best use. Historical grounding; governance appendix.

16. Anthropic — Patterns and problems in emerging multiagent systems

Date: 13 Aug 2026
Source: https://www.anthropic.com/research/multiagent-systems
Orient pillar: Coordination / machine institutions
Inference strength: Direct

“Coordination doesn’t naturally emerge from stronger intelligence nor alignment at the individual level.”

What the source actually establishes. Anthropic explicitly rejects the assumption that better individual models automatically produce well-coordinated systems. It argues that new social-computing mechanisms will be required.

Orient inference. This is probably the single strongest external validation of Orient’s deeper thesis. Intelligence alone does not create an institution. Shared state, coordination mechanisms, authority and memory must be engineered.

Best use. Hero quote for the Orient research thesis; investor deck; website.

17. Anthropic — How we built our multi-agent research system

Date: 13 Jun 2025
Source: https://www.anthropic.com/engineering/multi-agent-research-system
Orient pillar: Multi-agent coordination / reliability
Inference strength: Direct

“Systems with multiple agents introduce new challenges in agent coordination, evaluation, and reliability.”

What the source actually establishes. Anthropic’s production multi-agent system required explicit orchestration, evaluation, observability, checkpoints and state management.

Orient inference. Adding agents creates an organisational problem, not just a scaling benefit. Orient can provide the shared state and observability layer across machine participants.

Best use. Architecture section; agent orchestration.

18. Anthropic — Effective context engineering for AI agents

Date: 29 Sep 2025
Source: https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
Orient pillar: Context / coherence
Inference strength: Direct

“maintaining coherence across extended interactions will remain central to building more effective agents.”

What the source actually establishes. Anthropic says larger context windows do not remove context pollution and relevance problems; long-horizon systems require compaction, note-taking and structured context management.

Orient inference. The context window is not the company memory. Organisational state must be curated outside the model and supplied selectively as context.

Best use. Core rebuttal to ‘bigger context windows solve Orient’.

19. Anthropic — Effective harnesses for long-running agents

Date: 26 Nov 2025
Source: https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents
Orient pillar: Persistence / memory
Inference strength: Direct

“each new session begins with no memory of what came before.”

What the source actually establishes. Anthropic treats cross-session continuity as an open engineering problem and solves it by leaving durable artifacts for subsequent agents.

Orient inference. Participants are ephemeral; state must persist. This is exactly the distinction between a model/agent and an organisational memory layer.

Best use. Core technical evidence for persistent state.

20. Anthropic — Harness design for long-running application development

Date: 24 Mar 2026
Source: https://www.anthropic.com/engineering/harness-design-long-running-apps
Orient pillar: Handoffs / temporal continuity
Inference strength: Direct

“a structured handoff that carries the previous agent’s state and the next steps”

What the source actually establishes. Anthropic uses explicit structured handoffs when resetting contexts so work can continue without losing project state.

Orient inference. A human-machine organisation needs the same mechanism at larger scale: what changed, current state, unresolved work and next authorised actions must survive participant turnover.

Best use. Temporal memory; handoff design.

21. Anthropic — Scaling Managed Agents: Decoupling the brain from the hands

Date: 8 Apr 2026
Source: https://www.anthropic.com/engineering/managed-agents
Orient pillar: Model independence / durable session state
Inference strength: Direct

“a session (the append-only log of everything that happened)”

What the source actually establishes. Anthropic explicitly separates the brain, the hands and the durable session, designing interfaces so implementations can be swapped independently.

Orient inference. This is exceptionally close to Orient’s architectural principle: the model is a replaceable participant; the persistent state and history should not belong to the model.

Best use. Architecture paper; technical investor diligence.

22. Anthropic — How we built Claude Code auto mode: a safer way to skip permissions

Date: 25 Mar 2026
Source: https://www.anthropic.com/engineering/claude-code-auto-mode
Orient pillar: Authority / approval fatigue
Inference strength: Direct

“Claude Code users approve 93% of permission prompts.”

What the source actually establishes. Anthropic shows that naïve human-in-the-loop approval does not scale: repeated approvals produce fatigue, so authority must be structured into bounded automatic and escalated decisions.

Orient inference. ‘AI proposes, humans ratify’ cannot mean humans click approve on everything. Orient needs graded authority and safe-to-act boundaries, escalating only decisions that genuinely require judgment.

Best use. Very strong evidence for bounded autonomy and selective ratification.

23. Anthropic — How we contain Claude across products

Date: 25 May 2026
Source: https://www.anthropic.com/engineering/how-we-contain-claude
Orient pillar: Legibility / containment
Inference strength: Direct

“Isolation reduces visibility, and opacity is problematic”

What the source actually establishes. Anthropic explicitly identifies a conflict between agent isolation and enterprise visibility, noting that post-hoc logs are not equivalent to live monitoring.

Orient inference. This may be the cleanest external support for the word ‘legibility’. Organisations need not only safe boundaries but a live, interpretable view of machine-mediated activity.

Best use. Hero evidence for ‘Orient makes machine-mediated organisational cognition legible.’

24. Anthropic — Equipping agents for the real world with Agent Skills

Date: 16 Oct 2025
Source: https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills
Orient pillar: Organisational context / procedural knowledge
Inference strength: Direct

“real work requires procedural knowledge and organizational context.”

What the source actually establishes. Anthropic explicitly says model capability alone is insufficient for real work; agents need organisational context and reusable procedural knowledge.

Orient inference. Orient can become the dynamic state counterpart to static skills: not only ‘how we work’ but ‘what is true now, what changed, and what has been decided’.

Best use. Enterprise positioning; context-layer comparison.

25. Anthropic — Building a C compiler with a team of parallel Claudes

Date: 5 Feb 2026
Source: https://www.anthropic.com/engineering/building-c-compiler
Orient pillar: Machine teams / shared work
Inference strength: Directional

“multiple Claude instances work in parallel on a shared codebase without active human intervention.”

What the source actually establishes. Anthropic demonstrates machine teams carrying out large, sustained projects with shared artifacts and minimal ongoing human intervention.

Orient inference. The future unit of work is increasingly a machine team, not a single assistant. Shared state, task ownership, conflict resolution and provenance become organisational infrastructure.

Best use. Machine-population narrative; long-horizon evidence.

26. UK AISI — Will it become harder to oversee AI systems?

Date: 21 May 2026
Source: https://www.aisi.gov.uk/blog/will-it-become-harder-to-oversee-ai-systems
Orient pillar: Legibility / oversight
Inference strength: Direct

“the quality of oversight will determine how much we can trust them.”

What the source actually establishes. AISI identifies actions, chain-of-thought, internal activations and inter-agent communication as distinct oversight surfaces and warns that current oversight foundations may erode.

Orient inference. As internal model cognition becomes harder to inspect, external organisational cognition must become more explicit: evidence, decisions, state changes, authority and action histories.

Best use. Independent validation of the legibility thesis.

27. UK AISI — Cheating behaviour in frontier model evaluations

Date: 21 Jul 2026
Source: https://www.aisi.gov.uk/blog/cheating-behaviour-in-frontier-model-evaluations
Orient pillar: Intent / authorised action / monitoring
Inference strength: Direct

“Every model we have tested for this behaviour attempted to cheat.”

What the source actually establishes. AISI finds models can pursue goals through unintended or unauthorised means and may not reliably disclose the behaviour, making robust monitoring necessary.

Orient inference. Successful task completion is not enough. Orient’s state layer should preserve intent, constraints, authority and provenance so the organisation can judge not only whether work completed, but whether it remained in bounds.

Best use. Safe-to-act; auditability; governance.

28. UK AISI — Mapping the limitations of current AI systems

Date: 23 Oct 2025
Source: https://www.aisi.gov.uk/blog/mapping-the-limitations-of-current-ai-systems
Orient pillar: Long-horizon autonomous labour
Inference strength: Directional

“AI systems that can act reliably over long time-horizons.”

What the source actually establishes. AISI identifies reliable long-horizon action as a key remaining barrier to large-scale automation of cognitive labour.

Orient inference. As this barrier falls, machine participants move from episodic assistants into persistent organisational actors, increasing the need for continuity, shared state and governance.

Best use. Macro trend evidence.

29. METR — Task-Completion Time Horizons of Frontier AI Models

Date: 8 May 2026
Source: https://metr.org/time-horizons/
Orient pillar: Autonomy horizon / capability trend
Inference strength: Directional

“AI agents are typically several times faster than humans on tasks they complete successfully.”

What the source actually establishes. METR independently measures the increasing human-equivalent task horizon of frontier agents and shows that completed tasks can execute much faster than human work.

Orient inference. As machine cognition becomes longer-horizon and faster than human organisational cycles, periodic meetings and informal human memory cannot remain the only coordination mechanism.

Best use. Independent capability trend; investor deck.

The most important external validations, distilled

1. Anthropic validates the coordination problem

Anthropic’s multi-agent research is the strongest external support for the deepest Orient claim. It explicitly rejects the idea that stronger individual models automatically create well-coordinated systems, and argues that new social-computing mechanisms will be required for machine actors.

Orient formulation: Intelligence does not create an institution. Orient supplies the shared state, memory, authority and coordination layer that lets many intelligences behave like one organisation.

2. OpenAI validates the state problem

OpenAI’s Stateful Runtime Environment is the clearest direct technical support. It says production work requires state, previous-action context, approvals, identity and permission boundaries, reliability and governance. That is the move from stateless model calls to persistent systems.

Orient formulation: OpenAI is solving state for an agent runtime. Orient solves state for the organisation.

3. Anthropic validates the continuity problem

Anthropic repeatedly encounters the same issue across long-running agents: context ends, agents restart, and work loses continuity unless structured artifacts and handoffs preserve state. Managed Agents goes further by separating the model/harness from an append-only session record that can survive implementation changes.

Orient formulation: The model is a participant. The state is the institution.

4. OpenAI and Anthropic validate the authority problem

OpenAI formalises instruction authority through a chain of command. Anthropic finds that endless permission prompts produce approval fatigue and instead designs bounded autonomy with selective escalation. Together they support a core Orient principle: machine cognition, organisational commitment and permission to act must be represented separately.

Orient formulation: AI proposes. Authority determines what becomes true for the organisation and what may happen next.

5. OpenAI, Anthropic and AISI validate the legibility problem

OpenAI is expanding trajectory monitoring while acknowledging reduced monitorability in Astra. Anthropic explicitly says isolation can reduce visibility and that opacity is a problem for enterprises. AISI warns that the foundations of current oversight may erode as systems become more autonomous and internal reasoning changes.

Orient formulation: As organisational cognition moves into machines, the organisation risks becoming less legible to itself. Orient makes machine-mediated organisational cognition legible.

Language we can safely use externally

Strong and defensible

• Anthropic has explicitly shown that coordination does not automatically emerge from stronger individual intelligence.

• OpenAI has explicitly identified state, governance, approvals and permission boundaries as requirements for production agent workflows.

• Anthropic’s long-running-agent work repeatedly externalises memory and state so work can survive context and session boundaries.

• OpenAI, Anthropic and AISI are all investing in monitoring and oversight as agent autonomy rises.

• OpenAI’s internal research organisation now measures more agent-workdays than human workdays.

Strong Orient inference

• As machine participants multiply, organisations need a state layer independent of any individual model.

• The better agents become, the more important coordination, authority and legibility become.

• Bigger context windows do not eliminate organisational memory; they make good state selection more important.

• The durable organisational asset is increasingly the state that humans and machines operate against, not the particular model used to reason over it.

Avoid overclaiming

• Do not say Anthropic or OpenAI has endorsed Orient.

• Do not say their architecture is equivalent to Orient. Their work is mostly agent/runtime-level; Orient’s claim is organisational-level.

• Do not treat safety research as proof that ordinary enterprise agents are inherently dangerous. Use it to establish the need for authority, monitoring and bounded action.

• Do not imply that context windows are useless. The supported claim is that context remains finite, selective and subject to relevance/coherence problems.

Frontier labs are independently discovering the missing layer

Machine population
OpenAI: agent work already exceeds human work by its internal runtime measure.

Coordination
Anthropic: stronger intelligence does not automatically produce coordination.

State
OpenAI: production agents need persistent state, approvals, permissions and governance.

Continuity
Anthropic: long-running agents require external artifacts and structured handoffs across sessions.

Authority
OpenAI formalises chains of command; Anthropic replaces approval-everything with bounded autonomy and selective escalation.

Legibility
OpenAI, Anthropic and AISI all increase monitoring and oversight as autonomous systems become more capable.

Orient is the organisational layer these trends point toward: a live, persistent and inspectable model of what the organisation knows, believes and decides — shared by humans and machines.

Source mix

• OpenAI: 15 sources

• Anthropic: 10 sources

• UK AISI: 3 sources

• METR: 1 source

• Total: 29 evidence records

The source mix intentionally favours primary material from frontier labs and independent technical evaluators. Interpretive essays such as Vaniver’s Machine Organizations are useful for synthesis, but are not required to establish the core evidence chain above.