AI Engineer · Published 2026-10-08

Your AI Agent Has No Nervous System — Matt Gibiec, Dynatrace

privateReview: unreviewedConfidence: high
ObservabilityAI agentsopentelemetrydynatraceblueboxcontextagentsautonomy-levelscontext-engineeringruntime-topologysdlc
Open on YouTube ↗

Summary

Overview

  • Speaker: Matt Gibiec
  • Channel: AI Engineer
  • Main topic: AI agents lack operational context and observability, creating reliability and scaling challenges in production.
  • Purpose: To explain why AI agents fail or loop without production context, and to introduce Bluebox by Dynatrace as a solution for providing observability, context, and nervous-system capabilities to AI coding agents. Matt Gibiec, Regional Director of Solutions Engineering for AI-Native at Dynatrace, discusses the challenges developers, team leads, and heads of engineering face when deploying AI agents without observability, context, or runtime data. He introduces Bluebox by Dynatrace as an AI observability and context-provisioning tool designed to provide agents with production runtime context, automated service discovery, and evidence-based PR plans.

Topic Map

The Problem of AI Agents Operating Without Context

  • Explanation: AI agents are utilized to replace human tasks requiring reasoning, but they lack input, context, and a nervous system to understand their environment, leading to infinite loops and hallucinations.
  • Key claims:
    • AI agents operate at speed, but without sight.
    • Agents are only as good as the data available to make reasoning decisions.
    • Missing context and permissions lead to hallucinations and inefficient automated processes.
  • Examples:
    • An AI agent like Claude getting stuck in a loop repeating the same action and consuming tokens without producing desired results.
  • Terminology:
    • AI agent
    • hallucinations
    • reasoning
    • context
  • Why it matters: Without proper context, AI agents make decisions without understanding their consequences, wasting engineering resources and failing to deliver business value.

Challenges in AI-First Engineering Across Roles

  • Explanation: Different roles in engineering face distinct struggles when adopting AI agents, from developers dealing with runaway loops to heads of engineering wondering about ROI and code quality.
  • Key claims:
    • Developers struggle with why agents get stuck or fail to verify code independently.
    • Team leads struggle with PR deployments, stability, operational issues, and team focus.
    • Heads of engineering question if coding agents build the right things and deliver app value.
  • Examples:
    • Developers asking why an agent is stuck in a loop or cannot verify code on its own.
  • Terminology:
    • SDLC
    • operational issues
    • blast radius
    • PR plan
  • Why it matters: Organizations lack established instructions and metrics for measuring the value and success of AI agents.

Introducing Bluebox by Dynatrace

  • Explanation: Bluebox provides an agent integration layer across the development lifecycle to provide production context to AI coding agents and proprietary agents.
  • Key claims:
    • Bluebox scans repos, automatically discovers services and dependencies, and instruments with OpenTelemetry.
    • It builds live services topology and provides production runtime context to coding agents.
    • It helps find issues before they impact production.
  • Examples:
    • A feature working locally in development but breaking production services after being merged into the main branch.
  • Terminology:
    • Bluebox
    • OpenTelemetry
    • services and dependency mapping
    • runtime context
  • Why it matters: Providing production runtime context to coding agents prevents local changes from breaking downstream production services.

The Autonomy Ladder and Staying in Control

  • Explanation: AI automation requires an autonomy ladder ranging from observation to fully autonomous PR generation, allowing teams to maintain control.
  • Key claims:
    • Autonomy ladder allows teams to decide how far agents run without input.
    • Levels range from L1 (observe only), L2 (suggest fix), L3 (open PR), to L4 (autonomous PR).
    • Built on OpenTelemetry standards with GitHub PRs as output.
  • Examples:
    • Using L1 observation to build context before letting agents execute changes autonomously.
  • Terminology:
    • autonomy ladder
    • observe only
    • suggest fix
    • autonomous PR
  • Why it matters: Without control mechanisms, fully autonomous agents create a messy environment and erode trust in engineering output.

Key Points

AI agents lack a nervous system

  • Explanation: Agents execute tasks rapidly but lack real-time feedback from running software in development and production.
  • Evidence: Quotation from CTO Bernd Greifeneder: 'They operate at speed, but without sight.'
  • Practical implication: Engineering teams must provide continuous runtime context and observability to AI agents.

Context is King

  • Explanation: Context serves as the connective tissue that drives all automated processes and decision-making for AI agents.
  • Evidence: Without service dependency mapping and production data, agents make short-sighted decisions.
  • Practical implication: Agents need access to dependency graphs, historical loads, and application metrics before executing changes.

Local success does not equal production success

  • Explanation: Code tested successfully in local environments can break downstream services when merged into main branches.
  • Evidence: Merging code that passes local tests but breaks three other microservices.
  • Practical implication: Production context must be bridged into local testing and AI agent workflows.

Frameworks, Models & Processes

Bluebox Autonomy Ladder

  • How it works: A tiered framework defining the level of autonomy granted to AI agents during software delivery.
  • Components:
    • L1: Observe only
    • L2: Suggest fix
    • L3: Open PR
    • L4: Autonomous PR
  • When to use: When deciding how much trust and automation to grant AI coding agents in the SDLC.

Examples & Case Studies

A developer builds and tests a feature locally where it works perfectly, but merging it to the main branch breaks three other production services.

  • Illustrates: The lack of production dependency context in local AI coding agent workflows.
  • Lesson: AI agents need production topology and dependency mapping integrated into development and testing workflows.

Actionable Takeaways

  • Immediate:
    • Sign up for the Bluebox preview waitlist at bluebox.ai.
    • Evaluate whether your AI agents have access to sufficient runtime context.
  • Strategic:
    • Integrate observability and OpenTelemetry standards into AI agent workflows.
    • Adopt an autonomy ladder approach to maintain human control over AI-generated pull requests.
  • Questions to investigate:
    • How do we measure the business value and ROI of our AI agents?
    • How can we feed production runtime context directly into developer IDEs and coding agents?

Claims Worth Verifying

  • Bluebox provides evidence-based PR plans and automated service dependency mapping for AI agents. (Product Feature Claim)

Notable Quotes

"Your AI Agent has No Nervous System" (at 0:15) "They operate at speed, but without sight." (at 7:23) "Context is KING!" (at 10:37)

Compressed Summary

  • AI agents lack operational context and a nervous system to understand production environments.
  • Bluebox by Dynatrace integrates OpenTelemetry, service discovery, and topology mapping into AI workflows.
  • The autonomy ladder enables teams to scale agent automation safely from observation to autonomous PRs.
  • Keywords: observability, ai agents, opentelemetry, dynatrace, bluebox, context
  • Core insight: AI agents fail or introduce risk because they lack production observability and context, which tools like Bluebox provide by acting as a digital nervous system.

Core insights

6
Architecturemedium noveltystrong evidence

A graduated autonomy ladder (L1 observe-only, L2 suggest-fix, L3 open-PR, L4 autonomous-PR) treats trust in an agent as a per-capability setting rather than a binary deploy decision, letting teams accumulate L1 observation context before granting higher-level execution rights.

Why it matters

It places an explicit control boundary in the agent workflow where responsibility shifts from agent to human, and lets context-accumulation precede autonomy—so the same agent stack serves advisory and autonomous modes without redesign.

Generalization

For any code-modifying or action-taking agent, autonomy should be a tunable level with graduated intervention points, not a global on/off switch; lower levels double as data-collection phases for higher ones.

Levels range from L1 (observe only), L2 (suggest fix), L3 (open PR), to L4 (autonomous PR).
Open source video
Using L1 observation to build context before letting agents execute changes autonomously.
Open source video
Without control mechanisms, fully autonomous agents create a messy environment and erode trust in engineering output.
Open source video
Architecturemedium noveltymoderate evidence

Production runtime context—live service topology, dependencies, and load data—must be bridged backward into local development and AI coding-agent workflows, because locally passing code can break downstream services once merged.

Why it matters

It relocates observability from a post-deploy monitoring concern to a pre-merge input to the agent's reasoning loop, changing where a context/telemetry layer sits in the SDLC.

Generalization

Agents authoring changes to distributed systems need the runtime dependency graph as an input artifact before acting, not as post-incident diagnostics; otherwise local test success is a false signal.

Local success does not equal production success.
Open source video
A feature working locally in development but breaking production services after being merged into the main branch.
Open source video
Merging code that passes local tests but breaks three other microservices.
Open source video
Mechanismmedium noveltymoderate evidence

The same OpenTelemetry instrumentation that provides observability is repurposed as the context/memory substrate that agents consume—agents discover services by scanning repos and auto-instrumenting them, then read the resulting live topology.

Why it matters

It suggests telemetry data has a dual role: operational monitoring and agent-facing context, which changes the abstraction boundary for what feeds an agent's reasoning.

Generalization

Existing observability/instrumentation standards can serve as the context-provisioning layer for agents, avoiding a bespoke context pipeline and aligning agent context with production truth.

Bluebox scans repos, automatically discovers services and dependencies, and instruments with OpenTelemetry.
Open source video
It builds live services topology and provides production runtime context to coding agents.
Open source video
Failure Modemedium noveltymoderate evidence

A distinct failure mode is the runaway loop—an agent repeating the same action and burning tokens without producing results—attributed to missing environmental feedback rather than reasoning-model weakness.

Why it matters

It frames loop/stall as an observability-and-feedback problem in the environment, not just a prompting or model-quality problem, directing mitigation toward runtime feedback channels.

Generalization

Agent stalls/loops are frequently caused by the absence of a live feedback signal from the environment; adding runtime context can be more effective than tuning the model.

An AI agent like Claude getting stuck in a loop repeating the same action and consuming tokens without producing desired results.
Open source video
They operate at speed, but without sight.
Open source video
Practicemedium noveltymoderate evidence

Independent verification of code is an explicit agent capability gap: developers report agents that fail to verify code on their own, and production runtime context (evidence-based PR plans) is positioned as the substitute verification signal.

Why it matters

It reframes 'agent correctness' as an environment-instrumentation problem—giving the agent ground-truth evidence to check against—rather than a model capability problem.

Generalization

Verification loops for coding agents should be grounded in external runtime evidence (topology, metrics, load) rather than in the model's own self-assessment.

Developers struggle with why agents get stuck or fail to verify code independently.
Open source video
Bluebox provides evidence-based PR plans and automated service dependency mapping for AI agents.
Open source video
Mental Modelmedium noveltymoderate evidence

Agent value/ROI and code-quality questions are role-stratified across the SDLC—developers face loops and verification, team leads face PR stability and blast radius, heads of engineering face whether agents build the right things—implying no single success metric fits all levels.

Why it matters

It implies observability and evaluation must emit different signals for different consumers, so a single agent-metrics dashboard will not satisfy the SDLC.

Generalization

Agent operational instrumentation should be designed per-role/per-ownership level (blast radius, stability, business value) rather than as one flat metric set.

Team leads struggle with PR deployments, stability, operational issues, and team focus.
Open source video
Heads of engineering question if coding agents build the right things and deliver app value.
Open source video
Organizations lack established instructions and metrics for measuring the value and success of AI agents.
Open source video

Deep dives

5

Runtime evidence as the verification substrate for coding agents

Research question

Does grounding coding-agent PR generation and self-check in live production topology, dependency, and load evidence reduce downstream regressions compared with local-test-only or model-self-verification baselines?

Why

It tests whether agent correctness is primarily an environment-instrumentation problem rather than a model-capability problem, which determines where engineering investment should go.

Developers struggle with why agents get stuck or fail to verify code independently.
Open source video
It builds live services topology and provides production runtime context to coding agents.
Open source video
Bluebox provides an agent integration layer across the development lifecycle to provide production context to AI coding agents and proprietary agents.
Open source video
Source video

Empirical thresholds for a graduated autonomy ladder

Research question

At what autonomy level does added agent freedom stop improving merged-change outcomes and start degrading code quality, review burden, or reviewer trust?

Why

It converts a product control metaphor (L1-L4) into an empirically settable boundary, making autonomy a tunable per-capability parameter instead of a binary deployment decision.

Levels range from L1 (observe only), L2 (suggest fix), L3 (open PR), to L4 (autonomous PR).
Open source video
Using L1 observation to build context before letting agents execute changes autonomously.
Open source video
Without control mechanisms, fully autonomous agents create a messy environment and erode trust in engineering output.
Open source video
Source video

OpenTelemetry as the agent context/memory layer

Research question

Can existing OpenTelemetry instrumentation coverage and live topology serve as a sufficient, low-latency context substrate for coding agents, and where does uninstrumented-service blindness break that assumption?

Why

It determines whether teams need a bespoke context pipeline for agents or can reuse observability investments, directly changing the architecture of agent context provisioning.

Bluebox scans repos, automatically discovers services and dependencies, and instruments with OpenTelemetry.
Open source video
It builds live services topology and provides production runtime context to coding agents.
Open source video
Source video

Runaway loops as a missing-feedback failure mode

Research question

Do agent loops and stalls decrease more when runtime feedback channels are added than when model/prompt quality is improved?

Why

It reframes token-burning loops as an observability and environment-feedback problem, directing mitigation toward runtime signal injection rather than only model tuning.

An AI agent like Claude getting stuck in a loop repeating the same action and consuming tokens without producing desired results.
Open source video
AI agents operate at speed, but without sight.
Open source video
Source video

Role-stratified success metrics for agentic SDLC

Research question

Can per-role instrumentation signals (blast radius, PR stability, business value) predict AI-agent adoption value better than a single flat agent-metrics dashboard?

Why

It implies observability and evaluation must emit different signals for developers, team leads, and heads of engineering, so one metric set cannot serve the SDLC.

Team leads struggle with PR deployments, stability, operational issues, and team focus.
Open source video
Heads of engineering question if coding agents build the right things and deliver app value.
Open source video
Organizations lack established instructions and metrics for measuring the value and success of AI agents.
Open source video
Source video

Article ideas

4

Your Agent Doesn't Need a Better Model; It Needs a Nervous System

The dominant bottleneck in production coding agents is not reasoning capability but missing runtime feedback from the software they modify; instrumenting the environment reduces loops, hallucinations, and downstream breakage more effectively than model tuning.

Angle

Argue against model-centric agent improvement and for runtime-context engineering as the primary lever.

Source video

Autonomy Is a Ladder, Not a Switch

Treating autonomy as a per-capability graduated setting—observe, suggest, open PR, autonomous PR—lets teams accumulate context at lower levels before granting execution rights, preventing the trust erosion caused by premature full autonomy.

Angle

Use the L1-L4 control ladder to argue that agent trust should be accumulated, not deployed.

Source video

Local Tests Are a Lie for Distributed Systems

AI coding agents that rely on local test success as a correctness signal will merge changes that break downstream services; production dependency topology and load context must be pre-merge inputs to the agent, not post-incident diagnostics.

Angle

Show why local success is a false signal for agent-authored changes and what pre-merge runtime context changes.

Source video

Observability Is the New Context Layer for Agents

OpenTelemetry instrumentation should be treated as dual-purpose: operational monitoring for humans and machine-readable context/memory for agents, which means instrumentation coverage becomes a direct limit on agent quality.

Angle

Recast observability standards as agent infrastructure and discuss the blind spots created by uninstrumented services.

Source video

Project ideas

4

ContextBridge

movement-lab

Feeding a live OpenTelemetry service topology and dependency graph into a coding agent before it generates a PR reduces downstream production breakage per merged change compared with a local-tests-only baseline.

Proof of concept

Instrument a small microservice repository with OpenTelemetry, build a dependency/topology graph, expose it as an agent-readable pre-merge artifact to a coding agent (e.g., Claude/Cursor), and run a fixed set of feature changes.

Measurement

Downstream service breakage rate per merged PR, agent loop token cost, and reviewer-reported confidence versus a control group without topology context.

Source video

Autonomy Ladder Controller

gatehouse

A graduated L1-L4 autonomy gate with explicit control boundaries reduces code-quality degradation and reviewer trust erosion compared with full autonomous PR generation.

Proof of concept

Implement an agent workflow that can run at observe-only, suggest-fix, open-PR, and autonomous-PR levels on the same repository, with per-level approval gates and context-accumulation logging.

Measurement

Merged-change quality regressions, review burden (time per PR), token cost per merged change, and reviewer trust survey across matched L1-L4 runs.

Source video

Evidence-Grounded PR Verifier

beyond-evals

Coding-agent PR plans grounded in runtime evidence (topology, metrics, load) achieve higher PR acceptance and lower regression rates than plans grounded only in model self-assessment.

Proof of concept

Build a PR-plan generator that attaches live runtime evidence to each proposed change, then compare acceptance/regression outcomes against an identical agent without evidence grounding.

Measurement

PR acceptance rate, post-merge regression count, and false-positive fix rate across matched tasks.

Source video

Role-Signal Agent Metrics

new

Per-role agent success signals (developer loop/verification, team-lead blast radius/stability, head-of-engineering business value) predict adoption value and intervention needs better than one flat agent-metrics dashboard.

Proof of concept

Instrument an agentic SDLC pilot to emit role-specific metrics and compare their correlation with role-reported value and observed operational outcomes against a flat baseline dashboard.

Measurement

Correlation between role-specific metrics and expert ratings of value, plus changes in escalation/rollback rates after role-specific reporting.

Source video

Architectural implications

3

Observability is repositioned from post-deploy monitoring to a pre-merge input consumed by the agent.

Before

Telemetry primarily serves humans debugging production after deployment.

After

Telemetry is fed into the coding agent's context window before changes are proposed/merged.

Consequence

The context/telemetry layer becomes a dependency of the development loop, requiring low-latency, machine-readable topology and dependency data.

Source video

Agent control points are placed as graduated autonomy levels rather than a single human-approval gate.

Before

An agent either runs autonomously or is stopped for human review.

After

An explicit ladder (observe/suggest/PR/autonomous) governs how much the agent executes unreviewed.

Consequence

Workflow and PR tooling must support partial-autonomy handoffs and produce machine-generated PR artifacts as a first-class output.

Source video

Instrumentation standard (OpenTelemetry) is reused as the agent context abstraction.

Before

Context for agents is assembled ad hoc from prompts, files, and tool outputs.

After

Context is assembled from live instrumented topology and runtime data.

Consequence

Context quality becomes coupled to instrumentation coverage; uninstrumented services are blind spots for the agent.

Source video

Tradeoffs and failure modes

3

Autonomy vs. trust erosion

Benefit

Higher autonomy levels (L3/L4) let agents ship fixes with less human effort.

Cost or risk

Fully autonomous agents 'create a messy environment and erode trust in engineering output,' so moving up the ladder too fast can degrade code quality and confidence.

Without control mechanisms, fully autonomous agents create a messy environment and erode trust in engineering output.
Open source video
Source video

Context richness vs. loop/token cost

Benefit

Runtime context reduces hallucination and short-sighted decisions.

Cost or risk

Absent context, agents loop and consume tokens without results, making missing context a direct cost driver.

An AI agent like Claude getting stuck in a loop repeating the same action and consuming tokens without producing desired results.
Open source video
Source video

Local test success vs. production correctness

Benefit

Local development iteration is fast and self-contained.

Cost or risk

Relying on local success masks downstream breakage—passing local tests can still break three production microservices.

Local success does not equal production success.
Open source video
Source video

Open questions

4

How should production runtime context be fed directly into developer IDEs and coding agents?

Why unresolved

The summary raises this as a question to investigate and describes the capability at a product level without specifying the delivery mechanism or its latency/format constraints.

Research direction

Prototype a context-bridge that surfaces instrumented topology/metrics as an agent-readable artifact in the IDE/dev loop and measure change in downstream breakage.

Source video

How do we measure the business value and ROI of AI agents?

Why unresolved

The summary states organizations lack established metrics for agent value and success, and only poses the question.

Research direction

Define per-role success metrics (blast radius, PR stability, business value) and test whether an autonomy level correlates with measurable delivered value.

Source video

At what autonomy level does added agent freedom stop improving outcomes and start degrading quality/trust?

Why unresolved

The ladder is described as a control framework but the summary gives no empirical threshold or evidence for where to stop.

Research direction

Run matched experiments across L1–L4 on the same repo/workload and measure breakage, review burden, and token cost per merged change.

Source video

Is produced OpenTelemetry topology/dependency data sufficient to serve as the agent's verification signal, or is additional ground truth needed?

Why unresolved

The claim that runtime context substitutes for agent self-verification is asserted (evidence-based PR plans) but not demonstrated in the summary.

Research direction

Compare agent PR acceptance/regression rates with and without runtime-evidence grounding.

Source video

Key claims

7
causalVerification needed

AI agents execute at speed but lack real-time feedback from running software, i.e., they have no 'nervous system.'

Evidence

They operate at speed, but without sight.

Question

Does adding runtime feedback measurably reduce agent failure/loop rates versus model-only improvements?

Source video
causalVerification needed

Without service dependency mapping and production data, agents make short-sighted decisions requiring dependency graphs, historical loads, and application metrics before executing changes.

Evidence

Agents need access to dependency graphs, historical loads, and application metrics before executing changes.

Question

Does pre-execution access to dependency graphs and load data reduce change-induced production breakage?

Source video
causalVerification needed

Code that passes local tests can break downstream production services when merged to main.

Evidence

Merging code that passes local tests but breaks three other microservices.

Question

How frequently do locally-passing AI-generated changes break downstream services, and does topology-aware context reduce it?

Source video
factualVerification needed

An autonomy ladder (L1–L4) lets teams decide how far agents run without input while maintaining control.

Evidence

Autonomy ladder allows teams to decide how far agents run without input.

Question

Do teams using graduated autonomy report better quality/trust outcomes than those toggling full autonomy?

Source video
factualVerification needed

Bluebox provides evidence-based PR plans and automated service dependency mapping for AI agents.

Evidence

Bluebox provides evidence-based PR plans and automated service dependency mapping for AI agents.

Question

Does the generated mapping accurately reflect real service dependencies, and are the PR plans evidence-grounded in runtime data?

Source video
causalVerification needed

Fully autonomous agents without control mechanisms create a messy environment and erode trust in engineering output.

Evidence

Without control mechanisms, fully autonomous agents create a messy environment and erode trust in engineering output.

Question

Does the presence/absence of explicit control gates causally change code quality or reviewer trust metrics?

Source video
factualVerification needed

Bluebox instruments repos with OpenTelemetry and builds a live services topology to supply production runtime context to coding agents.

Evidence

Bluebox scans repos, automatically discovers services and dependencies, and instruments with OpenTelemetry.

Question

Does auto-instrumentation achieve sufficient coverage and accuracy to be a reliable agent context source?

Source video

Connections

4