Roadmap

Four phases. Pillar maturity over time.

Four phases for the agent runtime infrastructure for regulated production environments. Items move between tiers as the product ships — not as they are planned. The sequence is driven by what regulated enterprise buyers actually need next, not by what makes the feature list look complete.

Phase 1 — deployable today with design partners
Next — committed direction, in build
Later — sequenced, not speculative
Vision — direction, not commitment
Phase 01 · Design partners

Estate Foundation

Pillars 1–3 (Runtime, Authority and Accountability, and Lifecycle Control) shipped full. Pillar 4 (Audit-Grade Evidence) shipped foundational. Pillar 5 (Estate Intelligence) in development. Two framework adapters. Tier A deployment.

Registry & Identity
  • Agent registration and manifest — canonical inventory across every framework and team
  • Machine identity per agent — never user-borrowed credentials; every action attributable to a specific non-human principal
  • Framework adapters (two frameworks) — tool call interception and context propagation; narrow by design
Execution & Lifecycle
  • Durable execution engine — event-sourced, replayable, self-hosted; no vendor control plane dependency
  • Lifecycle authority — suspend / cancel / kill / drain / rollback at run, agent, version, or tenant scope
  • MCP gateway enforcement — single authorisation point for every tool call; paired with egress control
  • Sandboxed execution — process and network isolation per run; graduated profiles
Control & Observability
  • Token and resource limits — per-agent, per-tenant, per-run; enforced before consumption with circuit breakers
  • Cost attribution — per-run cost from registration metadata; no manual tagging
  • Tamper-evident audit log — hash-chained, privacy-preserving, compliance team query interface
  • Events journal — real-time operational stream with OTel export
  • Control Center — Platform Engineer and Security/Compliance views; full API surface
Deployment
  • Tier A — VPC / cloud-isolated — Helm chart deploy; EKS / AKS / GKE; no cogward control plane after deployment
  • Tool and datasource governance — the platform governs access; it does not build or maintain connectors

Phase 1 success condition: One regulated enterprise — financial services — running production agents on cogward. Pen test survived. Compliance audit referenced. That reference is the asset everything else is built on.

phase 02
Phase 02 · Next

Production Governance

Pillar 4 (Audit-Grade Evidence) reaches full — DORA / SOC 2 certification, GRC integrations. Pillar 5 (Estate Intelligence) foundation — task status, cost & latency baselines, goal achievement tracking. Evaluation gate, context governance, skills registry. Pillar 1 extended with Tier B deployment.

Governance
  • Policy hierarchy and pre-execution enforcement — permit, block, escalate before every significant action; never advisory
  • Policy versioning and approval workflow — full audit trail on policy changes; change approval before production promotion
  • HITL approval queue — ServiceNow, Jira, PagerDuty integrations
  • Production readiness checks — pre-promotion gate against policy and resource baselines
  • Compliance reporting templates — SOC 2, DORA, and NIST AI RMF as the first pack, driven by first customer requirements
  • GRC/SIEM integration depth — structured audit export to ServiceNow, Archer, Vanta, Splunk, and Sentinel
Evaluation & Evidence
  • Minimal production evaluation gate — behaviour validation, policy simulation against test traces, evidence-backed promotion decision
  • Forensic trace logging — capture of execution event loops, retrieval contexts, and local memory states during tool calls
  • GRC integrations — structured audit log exports to ServiceNow, Archer, and Vanta for real-time risk verification
Operations
  • Governed memory access layer — secure filtering abstraction over existing enterprise memory infrastructure with zero data migration required
  • Cost centre attribution and export — integration with SAP, Oracle, and internal financial chargeback engines
  • Agent Owner / Business view — run history logs, cost attribution trends, and human-in-the-loop escalation queues in Control Center
  • Tier B deployment envelope — fully on-premise execution architecture; bare-metal and self-managed Kubernetes targets with zero cogward control plane dependency

Phase 2 success condition: Multiple production customers. First compliance certification complete — one framework, one vertical, validated with real auditors. Governance layer generating real policy telemetry, not speculative configuration.

phase 03
Phase 03 · Later

Agent Lifecycle & Deployment Maturity

Pillar 5 (Estate Intelligence) reaches full — behavioural drift detection, fleet pattern intelligence, air-gapped distribution. Pillar 4 extended — snapshot-based forensic replay, data lineage. Pillar 1 extended — Tier C deployment. Sector compliance packs.

Goals & Outcomes
  • Goal declaration and outcome tracking — declaration, progress signals, outcome attribution; from "did this agent run" to "did this agent achieve its objective"
  • Task progress tracking — runtime progress visibility across long-running agent tasks
  • Runtime traces linked to evaluation outcomes — execution evidence tied directly to business outcome records
Evidence Depth
  • Snapshot-linked forensic replay engine — event history + retrieval context + memory state = deterministic replay input
  • Data lineage through agent execution — per-output chain from response to specific data source, access control, and retrieval timestamp
  • Policy sandbox over historical traces — test proposed policy changes against recorded production traces before going live
  • Sector-specific compliance packs — HIPAA, MiFID II, EU AI Act Article 13, following first customer validation
Lifecycle Integrations
  • Extended memory governance — cross-session agent memory parameters with structural residency controls and retention policies
  • Context provider registry — declarative, governed access to organizational context providers via the agent manifest schema
  • Organisational knowledge access governance — centralized access management for Confluence, SharePoint, and proprietary data stores without data duplication
  • Agent-to-agent governance — PII boundary enforcement, automated permission inheritance (deny-by-default), and multi-agent delegation audit chains
  • Air-gapped intelligence distribution pipelines — one-way delivery of pre-trained behavioral models and anomaly patterns pushed via signed physical media
  • Tier C deployment pipeline — air-gapped, sovereign, and highly classified defense estates; initialization of FedRAMP / IL-5 compliance engineering tracks

Phase 3 success condition: The platform is the operational system of record for the enterprise agent estate — not just the runtime for individual agents. Goals, outcomes, and lineage are visible and attributable. First defence reference customer in design partner programme.

phase 04
Phase 04 · Vision

Managed Options & Autonomous Operations

Full estate lifecycle system of record. Managed service offering (after self-hosted credibility). Governed self-improvement infrastructure. Standards contribution for manifest schema and audit event format.

Managed Options
  • Managed operations for BYOC — operational support, upgrade guidance, health review, and deployment hardening inside the customer environment
  • Dedicated managed control plane — optional cogward-operated control plane for organisations that accept managed infrastructure
  • Optional cloud-managed runtime — Tier A equivalent, vendor-operated, same compliance evidence as self-hosted
  • Partner-operated deployments — SI and partner-run cogward deployments for large enterprise programmes
Autonomous Operations
  • Long-running agent reference patterns — the durable execution engine already supports hours-long tasks; formalise as a first-class capability as the market matures toward day-long and week-long autonomous work
  • Governed self-improvement research track — self-improvement only when outcomes are verifiable and the audit evidence chain is mature enough to govern them safely
Ecosystem
  • Vertical evidence packs — additional sector-specific compliance packs as the customer base expands into new regulated verticals
  • Broader adapter ecosystem — community and commercial adapters for the broader framework ecosystem

Phase 4 direction: The managed service extends cogward's reach to the broader enterprise market without compromising the self-hosted story. Long-running agents and self-improvement are enabled by the platform's existing architecture — the right moment to formalise them as first-class capabilities is when the regulated enterprise base is proven and the evidence chain is mature enough to govern them safely.

Not building now

A proprietary organizational knowledge system. Enterprises have invested deeply in Confluence, SharePoint, and proprietary databases. cogward governs access to these systems — it does not build or replace them.
A generic connector marketplace. The platform enforces governed access to tools through the MCP gateway. Building and maintaining a connector catalog is a distribution play, not a runtime infrastructure play.
A generic skills marketplace. The agent manifest schema is the skills layer. A separate skills marketplace is a distribution play, not an infrastructure play.
A no-code agent authoring product. cogward wraps frameworks. It does not replace them or build a visual authoring surface.
A generic observability platform. The events journal exports to OTel-compatible SIEM and observability tooling. cogward does not compete with Datadog, Splunk, or general-purpose observability.
A model-provider marketplace. The platform governs how agents access models through the MCP gateway. It does not have preferred model relationships and does not bundle model access. This keeps cogward model-neutral permanently.
A self-improvement system before the evidence and evaluation chain is mature. Self-improvement only works reliably when outcomes are objectively verifiable. Building self-improvement infrastructure before the audit evidence chain is mature enough to verify what changed and why is unsafe and unprovable.

cogward governs how production agents access these systems. It does not need to own all of them.

Design partner programme

Co-design the production path.

Design partners get direct engineering access, early access to Phase 2 capabilities, and meaningful input into the roadmap sequence. The audit format, deployment runbooks, and compliance reporting templates are validated against your requirements — not hypothetical ones. This is not a waitlist. It is a working relationship.

What we need from you
A real deployment target — actual framework, actual environment, actual compliance requirement
What you get
Direct engineering access, Phase 2 early access, and roadmap input
Who we are looking for
Regulated financial services, healthcare, and defence teams deploying agents into production
Not a waitlist
We are taking a small number of design partners. Applications reviewed within 5 business days.
Apply for design partner programme