Agent/works is a governed agent runtime, OI's governance layer is a scoring taxonomy, so they overlap on vocabulary and not on function
The question
What is Thoughtworks Agent/works (launched 2026-06-16), and does its policy-enforcement and registry layer overlap the Organizational Intelligence governance layer?
Context: surfaced as a follow-up in [[2026-09-27-thoughtworks-agentic-scope-of-authority-nine-blocks]], which found the vault had zero coverage of Agent/works. This is an internal-platform comparison for the founder's phData role. It is not a sellable-differentiator question, since RDCO is dormant. Naming note: "CAF" was retired 2026-08-10; the comparison target is Organizational Intelligence (OI): Organizational Platform to Organizational Map, Intelligence Platform to Intelligence Maturity Assessment (IMA), plus the Organizational Intelligence Process (OIP) and its Pulse cadence.
What we already know (from the vault)
- Agent/works launched 2026-06-16 and both earlier Thoughtworks briefs missed it. The September brief logged five capability headings from the product page (Policy-First Enforcement, Built-In Enterprise Controls, Unified Registry & Runtime, Intelligent Memory Layer, Local Enterprise Control) and noted that AI/works "runs directly on the platform." [[2026-09-27-thoughtworks-agentic-scope-of-authority-nine-blocks]], [[2026-07-25-thoughtworks-aiworks-market-tier]], [[2026-07-15-agentic-assessment-framework-competitive-landscape]]
- OI's "Governance Layer" is one of six layers in a capability taxonomy, not software. The Intelligence Platform's Capability Registry is 6 layers, 28 sub-capabilities, 159 capability details. The Governance Layer (
GOV) has four sub-capabilities: Data & Information Governance, AI & Model Governance, Security/Privacy/Risk, and Strategic Oversight & Value Management. A client's capabilities are scored against it on a maturity model, and the output is the IMA. The registry is a list of things a client might have. It registers no agents and enforces nothing. (Source: OI spec working tree, read 2026-11-12; structure also cited in [[2026-09-06-move1-anchor-screen-oi-skill-catalog]].) - phData's runtime-side analogs sit in other teams' spokes, per a July map. The 2026-07-06 hub-and-spoke read has Registry ("the spine", a skill repo), Govern ("the control plane", IaC plus RBAC/IAM) and Evals ("the measurement plane") as separately owned spokes, with the founder's work as the Assess spoke. That map is four months old and was not re-verified for this brief. [[2026-07-06-die-fabric-hub-spoke-map-and-roadmap-implications]]
- The vault already has the runtime-versus-method distinction for a different vendor. Accenture's AI Refinery was classed as a runtime that does not contest OI's assessment-to-build ground. [[2026-09-11-accenture-ai-refinery-runtime-vs-methodology]]
- Every vendor is racing to be the agent registry of record. No neutral cross-substrate registry standard exists, and A2A declines to define a registry API. [[2026-09-13-cross-substrate-agent-identity-registry-standard]], [[2026-11-09-a2a-registry-discovery-signed-agent-cards]]
What the web says
Evidence tiers: (primary) = fetched from thoughtworks.com on 2026-11-12, read through a summarizing fetch tool, so exact wording outside quote marks is paraphrase. (search-summary) = appeared only in a search-engine summary and was not confirmed on a fetched page.
- What it is: "a single control plane and a governed runtime" for enterprise agents, "deployable on any cloud." The launch release (2026-06-16, San Francisco, timed to Databricks' Data + AI Summit) targets agent sprawl, risk and AI spend. A Databricks product director is quoted, and Thoughtworks says it is "partnering with Databricks." (launch release) (primary)
- Policy enforcement, as claimed in the release: before an agent runs, the platform analyzes every path through its workflow and confirms "at least one fully compliant path exists end to end"; permissions are "capability-based, scope-bound, and time-limited" and "narrow automatically" as the agent touches sensitive systems; scoped permissions carry across handoffs to third-party agents. These are vendor claims with no mechanism described. (launch release) (primary, marketing claim)
- Policy enforcement, as documented in the technical guide: a "Governance & Access" feature that defines "who can use which models, tools and platform resources," enforces access policies while agents work, and adds approval workflows for sensitive platform changes. That is access control over platform resources. The guide page as fetched does not describe the pre-execution path analysis, a policy language, or a policy engine. (technical guide) (primary)
- Registry: three inventories. The technical guide lists an Agent Registry ("agents, versions and deployments in one place," with promotion of tested versions between environments), a Model Catalog (cloud and self-hosted models, centralized credentials) and Tool Connections over MCP (tool discovery, per-agent tool access). The product page FAQ says it "registers agents, models, tools and data sources in a governed inventory." (technical guide, product page) (primary)
- The rest of the feature list: Monitoring & Audit (execution traces and audit records), Memory & Context (configurable persistence and sharing), Usage & Cost Controls (budgets, limits on model use, tool calls and task duration), Evaluation & Optimization (version comparison, prompt tuning). The product page names integration targets rather than its own engine: LangGraph, Bedrock Agents, Firecracker microVMs and Cloud Run for runtime; NeMo, Bedrock, Databricks and Vertex guardrails; OTEL, MLflow and CloudWatch for observability. (technical guide, product page) (primary)
- Deployment and commercial model: services-delivered, in the client's environment. The product page says Thoughtworks forward-deployed engineers configure it "entirely in your environment," and the only call to action is a sales conversation. No pricing, no GA or beta label, no licensing terms, no self-serve docs, and no named customer on any of the three pages. The only stated production user is Thoughtworks' own AI/works. A post-launch search found no independent review, analyst evaluation or customer case study. (primary for the pages; absence is as of 2026-11-12)
- Unverified implementation details. One search summary attributed to the technical guide said enforcement is Cerbos-based, agent definitions use Oracle's AgentSpec DSL, and agents carry full lineage tracking. None of that appeared on the technical guide as fetched, and a second targeted search found no source tying Cerbos or AgentSpec to Agent/works. A longer variant of the guide exists at a second URL that was not fetched (fetch cap reached). Treat all three details as unconfirmed. (search-summary, unverified)
- The five pillar names could not be reproduced. The product page as fetched today shows six differently worded capability blocks (one governance layer across the stack, audit trail, verified lower token spend, model choice, consistent governance as the stack evolves, runs wherever you do), and the technical guide shows eight features. Neither shows the five headings the September brief logged, and a search on the exact phrases returned nothing. Either Thoughtworks rewrote the page after 2026-09-27 or the fetch summarizer dropped them. Not resolved here.
Capability overlap table
OI column key: IMA taxonomy means the OI Capability Registry names the capability as something to score at a client. It does not mean OI provides it. "Evidence" is the tier for the Agent/works cell; the OI cell is from the spec working tree in every row.
| Agent/works component | What it does | Nearest OI component | Functional overlap | Evidence |
|---|---|---|---|---|
| Governance & Access (policy enforcement at run time) | Enforces who and which agent can use which models, tools, resources | GOV.AI_GOV.AGENT_GOV, GOV.SECURITY.ACCESS_CONTROL, INFRA.MLOPS.RUNTIME_GUARDRAILS (IMA taxonomy) |
None. OI scores whether the client has this; it enforces nothing | Primary |
| Pre-execution compliant-path analysis; narrowing time-limited permissions | Checks a workflow before it runs; scopes permissions per agent | GOV.AI_GOV.AGENT_GOV (IMA taxonomy) |
None as function; claim is itself unverified | Primary, marketing claim only |
| Approval workflows for sensitive changes | Human sign-off gate inside the platform | IFACE.HITL.REVIEW_APPROVE, GOV.VALUE_MGMT.STAGE_GATE (IMA taxonomy) |
None | Primary |
| Agent Registry | Inventory of deployed agents, versions, environments | INFRA.MLOPS.REGISTRIES (IMA taxonomy). OI's own registries hold capabilities, roles, values and engagement objects |
None on content. Same word, different thing registered | Primary |
| Model Catalog; MCP Tool Connections | Approved models and tools, with per-agent access | INTEL.FOUNDATION, INTEL.AGENTIC.TOOL_CATALOG (IMA taxonomy) |
None | Primary |
| Monitoring & Audit | Execution traces, audit records | GOV.SECURITY.AUDITABILITY, INFRA.ORCHESTRATION (IMA taxonomy) |
None | Primary |
| Memory & Context | Agent memory persistence and sharing | INTEL.AGENTIC.MEMORY (IMA taxonomy) |
None | Primary |
| Usage & Cost Controls; "verified lower token spend" against a baseline | Budgets and limits; savings claim checked against an agreed baseline | Business Value Framework, claimed-versus-realized per delivered Upgrade | Partial, weak. Both verify a value claim against a baseline; Agent/works covers token cost only, OI covers business value of a workflow change | Primary |
| Evaluation & Optimization | Compare agent versions, tune prompts | OIP Pulse (re-prioritize backlog as capabilities shift) | None. Pulse iterates on use cases and workflows, not on agent versions | Primary |
| Fleet-wide "single source of truth" | One view of agents, models, tools, policies | Engagement registry (Job, Workflow, Gap, Use Case, Upgrade) | Partial, adjacent. Both are a system of record that persists between rounds. Agent/works records what is running; OI records why it was built and what it changed | Primary |
| (nothing found) | Organizational Platform / Organizational Map (functions, motions, roles, seats) | None. No org-structure or role model on any fetched page | Primary, by absence | |
| (nothing found) | IMA maturity scoring; OIP Kickstart; use-case intake and prioritization | None. No assessment, maturity model or prioritization method on any fetched page | Primary, by absence | |
| Cerbos policy engine; AgentSpec agent definitions; lineage | Implementation specifics | n/a | Unknown | Search-summary, unverified |
Separately, and outside OI: against the July spoke map, Agent/works' agent registry, policy enforcement and evaluation features line up with the Registry, Govern and Evals spokes that other phData teams own. That is where a functional overlap would sit, if there is one. It is marked unknown here because the vault's picture of those spokes is from July and their current scope was not checked.
Convergences and contradictions
- Convergence: same pattern as Accenture. As with AI Refinery in [[2026-09-11-accenture-ai-refinery-runtime-vs-methodology]], the vendor's product is a runtime and control plane, and OI is a method with deliverables. The September brief's premise that Agent/works might be "a closer comparison set than AI/works" holds for the phData platform as a whole and does not hold for OI.
- Contradiction: the September brief's five pillar names are not on the pages as fetched today. See the last bullet under "What the web says." The September brief marked them primary-verified. The disagreement is unresolved.
- Contradiction inside Thoughtworks' own material. The release leads with pre-execution path analysis and auto-narrowing permissions. The technical guide documents role-and-resource access control with approval workflows. The stronger claims are in the press release only.
Synthesis for RDCO
The direct answer is no. Agent/works' policy-enforcement and registry layers do not cover the same ground as the OI governance layer, because the two are different kinds of object. Agent/works is deployed software that sits in the execution path of agents: it holds an inventory of agents, models and tools, and it allows or blocks actions. OI's Governance Layer is a section of a capability taxonomy that a client is scored against. The overlap is at the level of nouns. Nearly every Agent/works feature has a matching line in the OI Capability Registry, which is what you would expect, since the registry is meant to enumerate what a mature platform contains. Read that way, Agent/works is one possible implementation of roughly eight capability details that an IMA would score, spread across the Governance, Intelligence, Interface and Infrastructure layers. It is an answer to an IMA finding and does not compete with the instrument that produces the finding.
What Agent/works lacks is the half OI is built around. Nothing on the three fetched pages models an organization's functions, roles or workflows, scores maturity, or prioritizes use cases by business value. Its unit of record is the agent; OI's is the job and the workflow. The two real points of contact are narrow. Both keep a persistent system of record, and both verify a savings claim against a baseline, though Agent/works does so for token spend only. Neither is enough to call it a comparison set for OI.
The closer comparison is to the parts of the phData platform the founder does not own. If the July spoke map still holds, Agent/works bundles into one product what phData has as separate Registry, Govern and Evals spokes, and Thoughtworks sells it the same way phData would deliver those: forward-deployed engineers, in the client's environment, no public price. That is worth knowing for a conversation with those spoke owners, with two cautions. First, this brief did not verify the spokes' current state. Second, the evidence that Agent/works is more than a packaged services engagement is thin: no named customer, no independent review, no public documentation, and the only stated production user is Thoughtworks itself, five months after launch.
Confidence. High that there is no functional overlap with OI as specified, since that rests on the OI spec and on what Agent/works' own pages do and do not list. Medium on the feature inventory, which came through a summarizing fetch. Low on implementation details (policy engine, agent definition format) and on anything about adoption.
Why this is in the vault
It closes the Agent/works coverage gap opened by [[2026-09-27-thoughtworks-agentic-scope-of-authority-nine-blocks]] and settles that Agent/works does not belong in OI's comparison set, so OI roadmap and socialization material does not need to position against it. It also flags that the relevant comparison, if any, is against the Registry, Govern and Evals spokes in [[2026-07-06-die-fabric-hub-spoke-map-and-roadmap-implications]].
Open follow-ups
- Does the longer Agent/works technical guide name a policy engine, an agent definition format, or the pre-execution path analysis mechanism? A search summary said Cerbos and Oracle AgentSpec; the fetched guide page showed neither, and a second URL variant was not read. This decides whether the release's strongest claims are documented or marketing only.
- Has any enterprise publicly confirmed running Agent/works in production, or is AI/works still the only stated user? Five months after launch no customer is named on Thoughtworks' pages and no independent evaluation was found.
- Did Thoughtworks rewrite the Agent/works product page between 2026-09-27 and 2026-11-12? The five pillar headings logged in September are absent today. An archive comparison would show whether the positioning shifted from "policy-first enforcement" toward cost and model choice, or whether one of the two reads was wrong.
Related
- [[2026-09-27-thoughtworks-agentic-scope-of-authority-nine-blocks]] — parent brief; source of this question and of the five pillar names
- [[2026-07-25-thoughtworks-aiworks-market-tier]] — AI/works, which Thoughtworks says runs on Agent/works
- [[2026-07-15-agentic-assessment-framework-competitive-landscape]] — original competitive landscape; missed Agent/works
- [[2026-09-11-accenture-ai-refinery-runtime-vs-methodology]] — same runtime-versus-method verdict for Accenture
- [[2026-09-13-cross-substrate-agent-identity-registry-standard]] — vendors racing to be the registry of record
- [[2026-11-09-a2a-registry-discovery-signed-agent-cards]] — A2A defines no registry API
- [[2026-07-06-die-fabric-hub-spoke-map-and-roadmap-implications]] — Registry, Govern and Evals spokes
- [[2026-09-06-move1-anchor-screen-oi-skill-catalog]] — OI Capability Registry shape (6 / 28 / 159)
Sources
Vault
- [[2026-09-27-thoughtworks-agentic-scope-of-authority-nine-blocks]]
- [[2026-07-25-thoughtworks-aiworks-market-tier]]
- [[2026-07-15-agentic-assessment-framework-competitive-landscape]]
- [[2026-09-11-accenture-ai-refinery-runtime-vs-methodology]]
- [[2026-09-13-cross-substrate-agent-identity-registry-standard]]
- [[2026-11-09-a2a-registry-discovery-signed-agent-cards]]
- [[2026-07-06-die-fabric-hub-spoke-map-and-roadmap-implications]]
- [[2026-09-06-move1-anchor-screen-oi-skill-catalog]]
OI working tree (phData-internal, outside the vault, read 2026-11-12)
- OI spec: index, Intelligence Platform page, OIP page, A1 Capability Registry, A5 Engagement Registry Schema
Web (all fetched or searched 2026-11-12)
- Thoughtworks, "Thoughtworks Launches Agent/works to Govern and Run Enterprise AI Agents Across Any Cloud," 2026-06-16: https://www.thoughtworks.com/about-us/news/2026/thoughtworks-launches-agent-works
- Thoughtworks, Agent/works product page: https://www.thoughtworks.com/agent/works/
- Thoughtworks, "Agent/works key features" technical guide: https://www.thoughtworks.com/agent/works/technical-guide
- Not fetched, longer guide variant: https://www.thoughtworks.com/what-we-do/enterprise-ai/agent/works/technical-guide
- Trade coverage seen in search results only, not read: https://itbrief.co.uk/story/thoughtworks-launches-agent-works-for-ai-governance , https://enterprisedna.co/resources/news/thoughtworks-agentworks-ai-agent-governance-platform-2026