Skip to content

Instantly share code, notes, and snippets.

@jfarcand
Created April 18, 2026 14:42
Show Gist options
  • Select an option

  • Save jfarcand/bfe31239238f1435e2105614bfefca41 to your computer and use it in GitHub Desktop.

Select an option

Save jfarcand/bfe31239238f1435e2105614bfefca41 to your computer and use it in GitHub Desktop.
Atmosphere — AI Agent Foundation v0.8 (Dynatrace 2026 gap analysis + brainstorm, supersedes v0.7)

Atmosphere — AI Agent Foundation v0.8 (Dynatrace 2026 gap analysis + what to build next)

Supersedes the shipping pass in v0.7. v0.8 is the forward-looking brainstorming seed: where do Atmosphere's shipped primitives land against the real enterprise pain points documented in Dynatrace's The pulse of agentic AI in 2026 (N=919 enterprise leaders, Nov–Dec 2025, ≥$100M revenue), and what do we need to add to land the 2026-2027 enterprise story?


What the report says (one paragraph)

50% of surveyed enterprises have agentic projects in limited production, 44% broad departmental adoption, 23% enterprise-wide. Top three use cases: IT ops (70%), data processing (51%), cybersecurity (49%). 74% expect budgets to rise. Average current spend $2–5M per org; top ROI expected from IT ops (44%). Dynatrace's core thesis: observability must become the control plane — a deterministic-facts layer that anchors probabilistic model decisions across seven tiers (business → application → orchestration → agentic → model → RAG → infrastructure). 64% of orgs are building both autonomous and human-supervised agents; only 13% go pure-autonomous. 69% of agentic decisions are still verified by a human today.


Top production-readiness gates (what gates pilot → prod, page 11)

Criterion % Atmosphere status Commit or module of record
Security & data privacy 59% ✅ Strong Sandbox (Docker + NetworkPolicy), AgentIdentity.CredentialStore, DENY_ALL / REQUIRE_PRINCIPAL admin authorizers
Accuracy/reliability of outputs 55% 🟡 Partial AiGuardrail SPI, StructuredOutputCapturingSession, LlmJudge eval — no default drift detector
Adequate monitoring + controls 44% ✅ Strong OTel spans, Micrometer, typed TokenUsage, streaming-text budget, CoordinationJournal
Compliance with policies/regs 43% ✅ Strong AgentIdentity audit log, admin endpoints, AiGateway choke point
Business-objective alignment / ROI 41% ❌ Gap AiRequest.metadata() can carry keys but no convention, no standard correlations
Human oversight/review 40% ✅ Strong @RequiresApproval, ApprovalStrategy.virtualThread, CheckpointStore.fork(), PermissionMode.PLAN
Scalability & performance 38% ✅ Strong Virtual threads, Reactor Netty HTTP/3, WebTransport, backpressure invariants
Ethics / fairness / bias 30% ❌ Gap Guardrail SPI exists but no default PII / toxic / sensitive impls

Coverage: 6/8 strong, 1 partial, 2 gaps.


Top technical barriers to production (pages 12–13)

Barrier % Atmosphere status
Security / privacy / compliance 52% ✅ Strong
Managing + monitoring at scale 51% ✅ Strong
Rules for autonomous-vs-approval 45% Direct hit (PermissionMode, @RequiresApproval)
Skilled-staff shortage 44% — (docs / tutorials / CLI templates help)
Visibility into agent decisions 41% 🟡 OTel + CoordinationJournal; no flow viewer
Monitoring perf/cost/compliance 40% AiGateway + TokenUsage + AiMetrics
Coordinating troubleshooting 40% 🟡 Admin UI aggregates; no unified incident flow
Diagnosing unexpected behavior 39% ❌ No anomaly detector
Tracing downstream impacts 39% 🟡 Trace IDs propagate; no business-event linkage

Coverage: 5/9 strong, 3 partial, 1 gap.


Trust mechanisms in use today (page 15)

Mechanism % Atmosphere
Data quality checks 50% Guardrail SPI; no default DQ impl
Human review 47% @RequiresApproval + console tool-call cards
Drift / anomaly monitoring 41% ❌ No default
Security validations 39%
Consistency / regression tests 37% LlmJudge, StubAgentFleet, AbstractAgentRuntimeContractTest
Logs + traces 27%

Dynatrace's seven-layer control plane (page 26) — Atmosphere coverage

Layer Purpose Atmosphere status
Business Traceable outcomes ❌ No standard metadata keys
Application UX + tracing ✅ OTel + streaming session spans
Orchestration Advanced metrics & analysis @Coordinator + AgentFleet + CoordinationJournal
Agentic Agent-to-agent comms ProtocolBridge (InMemory + MCP + A2A + AG-UI + gRPC), AgentResumeHandle
Model / LLM Model integrity AiGateway, TokenUsage, cache hints, AiGuardrail
RAG / vector DB Data retrieval + semantic analysis atmosphere-rag (Spring AI + LangChain4j) + ContextProvider
Infrastructure Compute / GPU / network Out of scope (cede to observability vendors)

Coverage: 5/6 in-scope layers strong, 1 gap (Business), 1 deliberate out-of-scope.


What v0.5 → v0.7 shipped, re-read against the report

  • Barrier #45 (autonomous-vs-approval rules) + Gate #40 (human oversight)AgentIdentity.PermissionMode and @RequiresApproval annotated live in the personal-assistant sample.
  • Barrier #52 (security) + Gate #59 (security/data-privacy)Sandbox primitive with NetworkPolicy; proven live in the coding-agent sample where a Docker alpine container clones, executes, and is destroyed per request.
  • Barrier #40 (monitoring perf/cost)AiGateway admits every LLM call through a single choke point; default permissive impl shipped, applications install enforcing impls.
  • Resilience chapter ("absorb disruption, recover quickly") — AgentResumeHandle + RunRegistry + DurableSessionInterceptor; reattach via X-Atmosphere-Run-Id header.
  • Manual review of A2A flows (44% monitoring method)ProtocolBridge standardizes every dispatch path, including in-JVM, with traces.
  • OpenAI-compat tool-round-trip fix — broader enterprise endpoint coverage; matters for the 42% on-prem hosting finding and for every OpenAI-compatible proxy.

v0.8 brainstorming targets — six gaps, ranked by barriers hit

1. Business-outcome metadata convention (business.* keys)

Why — Gates #41 (ROI alignment) and barriers #39 (downstream-impact tracing) and #39 (linking to business outcomes). The report is emphatic that enterprise buyers can't close the "AI → dollars" loop.

Shape idea — add a sealed interface / record on AiRequest.metadata() that documents a stable namespace:

  • business.customer.id
  • business.session.revenue (pre-call budget) / business.session.cost (post-call actual)
  • business.event (kind enum: new_conversation | purchase | support_escalation | churn_risk)
  • business.tenant.id

Downstream observability vendors (Dynatrace, Datadog, New Relic) can then correlate traces to revenue / CSAT / conversion out of the box.

Brainstorming questions

  • Do we ship this as a new module atmosphere-business-context, or fold it into atmosphere-ai as a constants class?
  • Stable OTel semantic conventions namespace (e.g., atmosphere.business.*) or reuse GenAI conventions?
  • Do we emit these as events or spans — or both (event attached to the active span)?

2. Default drift / anomaly AiGuardrail impls

Why — 41% of orgs do drift monitoring today as a validation; 39% cite "diagnosing unexpected behavior" as a top barrier. Atmosphere has the AiGuardrail SPI but ships no defaults.

Shape idea — three ship-with-it guardrails behind feature flags:

  • RollingEmbeddingDistanceGuardrail (per-agent rolling window; fire when response embedding is >N std-dev from the mean)
  • OutputLengthZScoreGuardrail (detect truncation or runaway generations)
  • ToolCallRateGuardrail (detect loops and tool-call flooding)

Brainstorming questions

  • Do we bring our own embedder, or require the app to configure one?
  • Where does the rolling window state live — AgentState or a dedicated DriftStore SPI?
  • How do we surface drift alerts? New AiEvent.DriftDetected? Metrics counter? Both?

3. Agent-to-agent flow viewer in the admin control plane

Why — 44% still review agent-to-agent communication flows manually (page 22). We already have the underlying data via ProtocolBridge events + CoordinationJournal; we just don't visualize it.

Shape idea — extend the admin React UI at /atmosphere/admin/ with a live flow graph: nodes = agents, edges = dispatches (labeled with skill, duration, success). Drill-down opens the CoordinationJournal entry + trace spans. Data source: WebSocket event stream + new REST endpoint /atmosphere/admin/flows.

Brainstorming questions

  • How far back do we show history — last 10 runs, last 1h, configurable?
  • Do we make this opt-in per user (privacy on multi-tenant admin deployments)?
  • Render library? (We already use React; Dagre / ELK / d3?)

4. Deterministic fact context SPI

Why — Dynatrace's core thesis (page 26): "AI actions must be grounded in deterministic, real-time facts." Atmosphere's ContextProvider is RAG-oriented (retrieve + rerank); we don't have a primitive for injecting verified facts (user profile, locale, time, feature flags, recent user actions).

Shape idea — new primitive FactResolver with per-key resolvers: user.profile, user.locale, time.now, time.timezone, featureflag.<name>, audit.last_n_actions. Results flow into AgentExecutionContext.facts(), captured per turn for replay/audit.

Brainstorming questions

  • Do we call this FactResolver, DeterministicContext, or GroundingContext?
  • Is it a ninth primitive or an extension of AgentState?
  • Auth model — how do we keep a fact resolver from leaking data across tenants?
  • Cache policy — facts usually change slowly; TTL per resolver?

5. Default ethics / PII / toxic-content guardrails

Why — 30% of orgs gate on ethics/fairness/bias today. Guardrail SPI exists; no defaults.

Shape idea — three implementations in atmosphere-guardrails-default:

  • PiiRedactionGuardrail (Microsoft Presidio-style patterns + config-driven custom regex)
  • ToxicContentGuardrail (small classifier — local model or remote API behind an SPI)
  • SensitiveTopicGuardrail (allowlist / denylist + LLM judge fallback)

Brainstorming questions

  • Zero-deps implementations (regex-only) vs model-backed?
  • Default GuardrailResultBlock or Modify (redact in place)?
  • How do we test these without shipping personal data? — synthetic fixtures.

6. "Intentional & enduring" human-in-the-loop tutorial

Why — Chapter 3 of the report calls out the three-phase trust build (preventive → supervised → full autonomous). Atmosphere has all the primitives; our docs don't narrate the incremental pattern.

Shape idea — new tutorial page on atmosphere.github.io: "Earning trust in autonomous operations — the three-phase adoption pattern." Walk through the same coordinator code with PermissionMode.PLANaccept-editsdefault; each phase links the guardrails and observability that make the next step safe.

Brainstorming questions

  • Where in the tutorial tree (between 09-ai-endpoint and 10-coordinator, or its own chapter)?
  • Do we need a companion sample, or can pa + coding-agent demonstrate the phases with config-only toggles?

Non-goals (proposed for v0.8)

To keep the roadmap honest:

  1. Business-metric dashboards — cede to observability vendors. We emit the data; they render.
  2. General-purpose model hosting — no inference, no GPU infra. Stay runtime-agnostic.
  3. Vendor-specific certifications (SOC2 etc.) — repo-level, not framework-level.
  4. Ethics model training — we ship SPIs + one default; app teams own policy.

Open brainstorming questions

  1. Do we cut atmosphere-ai 4.0.39 first (shipping the foundation) and line these up as 4.1.x, or bundle #1 (business-metadata) into 4.0.39 since it's an additive contract?
  2. Is there a natural order — e.g., #4 (fact-context) before #2 (drift) because fact-grounding changes what "drift" means?
  3. The report's Chapter 3 insight ("64% build both") — does it change our default PermissionMode from default (auto-approve) to PLAN (approval required)? Safer default, breaking change.
  4. Should the admin flow viewer (#3) become the anchor of a future "Atmosphere as the agentic control plane" narrative, positioning Atmosphere adjacent to observability vendors rather than competing?
  5. Of the six gaps, which two does ChefFamille want landed before the first public v0.7/v0.8 announcement, and which can slot into 4.1?
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment