AI Engineer · Published 2026-09-09

Your agents lack context: Here's how to fix "You're absolutely right!" — Brandon Waselnuk, Unblocked

privateReview: unreviewedConfidence: high
Context engineeringAI agentsRAGMCPtokensDeveloper productivitycontext-engineeringagent-architecturetoken-efficiencyretrievalenterprise-aipermissions
Open on YouTube ↗

Summary

Overview

  • Speaker: Brandon Waselnuk
  • Channel: AI Engineer
  • Main topic: Context engineering for AI coding agents and teams
  • Purpose: To educate AI engineers on why AI agents fail due to lack of organizational context and how to build or adopt a context engine that integrates data sources, resolves conflicts, and optimizes tokens. Brandon Waselnuk from Unblocked discusses context engineering for AI agents and how providing correct enterprise context solves the problem of AI models blindly agreeing with users ('You're absolutely right!'). He breaks down the AI maturity curve from tab complete to autonomous agent teams, outlines the failures of curated-context traps and MCP plateaus, details the six core requirements for a real context engine, and presents real-world impact metrics.

Topic Map

Introduction and Context Engineering

  • Explanation: Introduction to Brandon Waselnuk and the premise of context engineering to fix the tendency of AI models to say 'You're absolutely right!'
  • Key claims:
    • AI-generated code should feel like it was written by someone who's been on your team for years.
    • For years, humans have been the context engine.
    • As agents scale in autonomy, the cost of bad context compounds.
  • Examples:
    • Unblocked giving out coconuts for fresh context.
  • Terminology:
    • Context engineering
    • Tab complete
    • Doom loops
  • Why it matters: Without context engineering, AI agents waste tokens and time in doom loops and fail to align with internal team practices.

Levels of AI Maturity and the Cost of Bad Context

  • Explanation: Examining the adoption curve of AI agents from tab complete up to agent teams and how compounding context errors increase review tax and silent failures.
  • Key claims:
    • The cost of bad context compounds with every step toward autonomy.
    • Tab complete catches errors early, but agent IDEs lead to doom loops.
    • Parallel agents hit review tax where senior engineers spend the day reviewing instead of building.
    • Without human in the loop, silent failure occurs where code compiles but breaks production.
  • Examples:
    • An AI agent writing code that compiles but breaks production at 1 AM because it missed feature flag requirements.
  • Terminology:
    • Doom loops
    • Review tax
    • Silent failure
    • Background agents
  • Why it matters: Moving toward agentic autonomy without robust context engines multiplies organizational costs and risks silent failures.

Common Approaches That Don't Work

  • Explanation: Critique of common local maxima in context management: curated-context traps and MCP plateaus.
  • Key claims:
    • Curated-context repos rot faster than you can update them.
    • MCP plateau leads to satisfaction of search bias where agents stop at the first piece of information they find.
    • Access to information does not equal understanding.
  • Examples:
    • Markdown files in a virtual file system that quickly become outdated tribal knowledge.
    • Slack conversations from last night contradicting architecture records.
  • Terminology:
    • Curated-context trap
    • MCP plateau
    • Satisfaction of search bias
    • Local maxima
  • Why it matters: Static markdown files and unstructured MCP tool calls fail to capture the relational and dynamic nature of enterprise engineering.

How a Context Engine Works: The Six Requirements

  • Explanation: Detailed breakdown of the architecture and six key requirements of a true context engine.
  • Key claims:
    • A context engine ingests signals from code, docs, Slack, conversations, and incident management.
    • It requires unified system context, targeted retrieval, conflict resolution, personalized relevance, token optimization, and permission enforcement.
  • Examples:
    • Resolving conflicts between an old architecture diagram and last night's Slack convo with the CTO.
  • Terminology:
    • Unified system context
    • Targeted retrieval
    • Conflict resolution
    • Personalized relevance
    • Token optimization
    • Permission enforcement
  • Why it matters: Enterprise AI requires systems that understand user permissions, resolve conflicting data sources, and optimize token usage.

Impact and Results with Context Engines

  • Explanation: Benchmarking results comparing AI prompts with and without context engines.
  • Key claims:
    • Using a context engine reduced task time from 2 hours 33 minutes to 25 minutes (2 hours saved).
    • Token usage dropped from 20.9 million to 10.8 million tokens (48% fewer tokens).
    • Answer quality improves because answers are not buried in noise.
  • Examples:
    • Cloudbeds case study showing 50% fewer tokens, faster triage, and better answer quality.
  • Terminology:
    • Wall clock time savings
    • Token savings
    • Hydrated context
  • Why it matters: Hydrating agents with context drastically cuts developer time and API token waste.

Open Source Software and Tools

  • Explanation: Presentation of open source tools like Social Comment Network (Resheen the machine) and Repo Rules Agent.
  • Key claims:
    • Open source tools help analyze team interaction graphs and rule files across repositories.
    • Schema-less lookup allows agents to query schemas deterministically.
  • Examples:
    • Social Comment Network QR code and Repo Rules Agent QR code.
  • Terminology:
    • Social Comment Network
    • Repo rules agent
    • Schema-less lookup
  • Why it matters: Provides engineers with concrete open-source tooling to implement context engineering in their own organizations.

Key Points

The gap is not intelligence, it's context

  • Explanation: Models are increasingly intelligent (like Mythos or future models), but fail when they lack specific company and team context.
  • Evidence: Quoting Andrej Karpathy: 'The gap is not intelligence. It's context.'
  • Practical implication: Engineering effort should focus on context engineering rather than waiting solely for smarter base models.

Access to information does not equal understanding

  • Explanation: Giving agents access to all repositories and documents via naive RAG or MCP does not mean they understand organizational priorities, permissions, or temporal relevance.
  • Evidence: Comparison of underwater iceberg of code that compiles vs. hidden tribal knowledge like feature flag rollouts.
  • Practical implication: Context engines must resolve conflicts, apply recency signals, and respect governance.

Token and Time Efficiency via Context Hydration

  • Explanation: Hydrating prompts with pre-filtered, relevant context eliminates wasteful search loops and reduces token consumption.
  • Evidence: Cloudbeds benchmark: 2 hours 33 minutes down to 25 minutes; 20.9M tokens down to 10.8M tokens.
  • Practical implication: Context engineering directly lowers API costs and accelerates developer velocity.

Frameworks, Models & Processes

Levels of AI Maturity in Engineering Organizations

  • How it works: A staged adoption model tracking increasing agent autonomy from human-in-the-loop to fully autonomous teams.
  • Components:
      1. Tab complete (You are the context)
      1. Agent IDE
      1. Context engineering
      1. Compounding engineering
      1. MCP + skills (Context layer)
      1. Harness engineering
      1. Background agents
      1. Agent teams
  • When to use: To assess an engineering organization's readiness and maturity in deploying AI developer tools.

Context Engine Architecture (Six Requirements)

  • How it works: Ingests multi-source signals, processes them through six core constraints, and outputs optimized context to workflows.
  • Components:
    • Unified system context
    • Targeted retrieval
    • Conflict resolution
    • Personalized relevance
    • Token optimization
    • Permission enforcement
  • When to use: When designing or selecting backend infrastructure for enterprise AI agents.

Examples & Case Studies

An AI coding agent generates code that successfully compiles but breaks production because it missed a specific feature flag rollout rule.

  • Illustrates: The danger of code that compiles versus code that understands organizational intent and historical processes.
  • Lesson: Agents need deep organizational context beyond raw syntax and compiler feedback.

Cloudbeds used a context engine for support and engineering workflows.

  • Illustrates: Real-world productivity and cost benefits of context engineering.
  • Lesson: Achieved ~50% fewer tokens, faster triage, and higher answer quality.

Actionable Takeaways

  • Immediate:
    • Evaluate your team's current AI maturity stage at readiness.getunblocked.com
    • Explore open-source context tools for mapping repo rules and team comment networks
  • Strategic:
    • Shift focus from prompt engineering to context engineering
    • Implement context engines that resolve multi-source conflicts and enforce governance before delivery to agents
  • Questions to investigate:
    • How are conflicting instructions across Slack, Jira, and GitHub resolved in your current AI workflows?
    • Are your AI agents hitting satisfaction of search bias or token-wasting doom loops?

Claims Worth Verifying

  • Cloudbeds reduced task token usage by 48% and saved 2 hours of time using a context engine (empirical case study metric)
  • Andrej Karpathy stated: 'The gap is not intelligence. It's context.' (quote attribution)

Notable Quotes

"The gap is not intelligence. It's context." (at 21:50)

Compressed Summary

  • AI agents fail without proper organizational context, resulting in doom loops and review tax.
  • Curated markdown repos rot and MCP tools suffer from search bias.
  • A true context engine requires unified signals, conflict resolution, targeted retrieval, personalized relevance, token optimization, and permission enforcement.
  • Context engineering cuts token usage by ~50% and dramatically reduces developer resolution time.
  • Keywords: context engineering, AI agents, RAG, MCP, tokens, developer productivity
  • Core insight: As AI agent autonomy scales, the gap between model intelligence and successful execution is bridged not by better prompts, but by comprehensive context engineering.

Core insights

6
Architecturehigh noveltymoderate evidence

A production context engine is defined by six co-equal requirements — unified system context across code/docs/Slack/conversations/incidents, targeted retrieval, conflict resolution, personalized relevance, token optimization, and permission enforcement — rather than retrieval quality alone.

Why it matters

Most teams build a retrieval layer (RAG or MCP) and treat conflict resolution, personalization, and permissions as afterthoughts. Naming them as first-class requirements changes what components must be built and where the boundary between 'tool' and 'engine' sits.

Generalization

Any agent operating over multi-source enterprise data needs conflict resolution, recency, per-user relevance, and ACL enforcement as distinct subsystems, not as prompt instructions.

A context engine ingests signals from code, docs, Slack, conversations, and incident management.
Open source video
It requires unified system context, targeted retrieval, conflict resolution, personalized relevance, token optimization, and permission enforcement.
Open source video
Failure Modehigh noveltymoderate evidence

'Satisfaction of search bias' is a concrete agent failure mode: when retrieval is exposed as raw MCP tools, the agent stops at the first plausible piece of information rather than continuing to search for the authoritative source.

Why it matters

This reframes MCP as a local maximum. If the agent decides when to stop searching, reliability is bounded by the agent's own stopping heuristic; pre-filtering and ranking must move into the engine.

Generalization

Giving an agent unconstrained search tools pushes stopping-criteria decisions into the model, where first-match heuristics dominate over coverage or authority.

MCP plateau leads to satisfaction of search bias where agents stop at the first piece of information they find.
Open source video
Access to information does not equal understanding.
Open source video
Failure Modemedium noveltymoderate evidence

Hand-curated context repositories (markdown files in a virtual filesystem) rot faster than they can be maintained, becoming stale tribal knowledge that directly contradicts live sources such as Slack.

Why it matters

This challenges the common 'just write a CLAUDE.md / rules file' practice. Static curated context creates a floor of correctness that decays over time and introduces contradictions the agent has no mechanism to detect.

Generalization

Manually curated context is a perishable asset; any context source without an ingestion/refresh pipeline should be treated as degrading infrastructure.

Curated-context repos rot faster than you can update them.
Open source video
Markdown files in a virtual file system that quickly become outdated tribal knowledge.
Open source video
Slack conversations from last night contradicting architecture records.
Open source video
Mental Modelmedium noveltymoderate evidence

The cost of bad context compounds monotonically along the autonomy curve: tab complete surfaces errors early, agent IDEs create doom loops, parallel agents create 'review tax' on senior engineers, and fully autonomous background agents produce silent failures where code compiles but breaks production.

Why it matters

It gives a staged maturity model where each increment of autonomy raises the required context quality rather than just changing tooling. Teams scaling autonomy without context investment shift cost from authoring to review and then to incident response.

Generalization

Reliability requirements for context scale super-linearly with agent autonomy; a context investment that is adequate for human-in-the-loop is insufficient for unsupervised agents.

The cost of bad context compounds with every step toward autonomy.
Open source video
Parallel agents hit review tax where senior engineers spend the day reviewing instead of building.
Open source video
An AI agent writing code that compiles but breaks production at 1 AM because it missed feature flag requirements.
Open source video
Empirical Resultmedium noveltymoderate evidence

Token optimization via pre-hydrated context improves answer quality and cuts token usage simultaneously — the reported benchmark moved a task from 2h33m to 25m and tokens from 20.9M to 10.8M.

Why it matters

Directly contradicts the 'give the model more context / bigger window' intuition: more retrieved material means answers buried in noise, more search loops, and higher cost. This is a measurable, testable cost/quality lever.

Generalization

Pre-filtering and ranking context before it reaches the model can reduce cost and improve quality in the same intervention, because irrelevant context is actively harmful, not merely wasteful.

Using a context engine reduced task time from 2 hours 33 minutes to 25 minutes (2 hours saved).
Open source video
Token usage dropped from 20.9 million to 10.8 million tokens (48% fewer tokens).
Open source video
Answer quality improves because answers are not buried in noise.
Open source video
Mental Modelmedium noveltyweak evidence

Karpathy's framing — the gap is not intelligence but context — implies the primary engineering investment should shift from waiting for stronger base models to building context infrastructure.

Why it matters

It sets an investment priority: capability gaps observed in deployed agents are more likely context-acquisition failures than model-capability failures, so benchmarks and debugging should target context provenance first.

Generalization

When agents underperform, the first diagnostic should be what relevant information never reached the model, not which model was used.

The gap is not intelligence. It's context.
Open source video
Models are increasingly intelligent (like Mythos or future models), but fail when they lack specific company and team context.
Open source video

Deep dives

5

Deterministic conflict resolution across multi-source enterprise context

Research question

What resolution policy (recency weighting, source authority, human adjudication, or a hybrid) maximizes agent answer accuracy when an old architecture record contradicts a recent Slack conversation with an executive?

Why

Conflict resolution is named as one of six co-equal requirements of a context engine but no policy is described, so every team building one must invent it; without a tested policy the engine cannot be trusted at higher autonomy levels where silent failure is the cost.

It requires unified system context, targeted retrieval, conflict resolution, personalized relevance, token optimization, and permission enforcement.
Open source video
Slack conversations from last night contradicting architecture records.
Open source video
Source video

Measuring satisfaction of search bias and the MCP stopping heuristic

Research question

How often do agents terminate retrieval early when an authoritative source exists later in the search space, and does moving ranking/stopping into the engine measurably reduce that rate?

Why

It reframes MCP as a local maximum and locates reliability in the agent's own stopping heuristic; if the failure rate is high, retrieval control belongs in the engine, which is a concrete component-boundary decision.

MCP plateau leads to satisfaction of search bias where agents stop at the first piece of information they find.
Open source video
Access to information does not equal understanding.
Open source video
Source video

Detecting silent failures from missing institutional context

Research question

Can golden-task evals over institutional rules (feature-flag requirements, rollout policy) detect violations that compile correctly and never appear in the prompt, and how much does context hydration reduce them?

Why

Silent failure is the terminal cost on the autonomy curve, and no detection mechanism, eval, or monitoring approach is given; a detection method is the missing prerequisite for running unsupervised background agents.

An AI agent writing code that compiles but breaks production at 1 AM because it missed feature flag requirements.
Open source video
The cost of bad context compounds with every step toward autonomy.
Open source video
Source video

Staleness half-life of curated markdown context versus a maintained engine

Research question

What is the observed decay rate of hand-curated context repositories in active codebases, and does an ingestion/refresh pipeline avoid the same rot it criticizes?

Why

The curated-context trap and the engine are both maintenance-cost propositions; without measuring decay on both sides, the criticism of curated repos is asymmetric and the engine's own operational burden goes unquantified.

Curated-context repos rot faster than you can update them.
Open source video
Markdown files in a virtual file system that quickly become outdated tribal knowledge.
Open source video
Source video

Reproducibility of the reported context-engine gains

Research question

Do the 2h33m-to-25m task time and 20.9M-to-10.8M token figures hold across multiple repos and task types under a controlled no-engine baseline, or are they explained by task selection?

Why

These are the only quantitative claims supporting the whole context-engine argument; if they are single-case artifacts, the investment priority thesis rests on weak footing, so a controlled replication is the direct test.

Using a context engine reduced task time from 2 hours 33 minutes to 25 minutes (2 hours saved).
Open source video
Token usage dropped from 20.9 million to 10.8 million tokens (48% fewer tokens).
Open source video
Answer quality improves because answers are not buried in noise.
Open source video
Source video

Article ideas

4

Six requirements, not one retrieval layer: why most context stacks are only 1/6 of an engine

Teams that ship RAG or MCP have built the retrieval requirement and mistaken it for the whole product; conflict resolution, personalized relevance, and permission enforcement are separate subsystems whose absence, not retrieval quality, is what breaks autonomous agents.

Angle

Architecture teardown that maps each of the six requirements to a missing component in a typical RAG/MCP stack and shows where the tool-versus-engine boundary actually sits.

Source video

Your agent stopped searching too early: satisfaction of search bias is the real MCP ceiling

Exposing raw retrieval tools pushes the stopping decision into the model, where first-match heuristics beat coverage and authority; moving pre-filtering and ranking into the engine is a reliability change, not an optimization, and it is measurable.

Angle

Failure-mode analysis with a proposed eval harness that confronts the agent with an early plausible source and a later authoritative one.

Source video

Context engineering is a perishable-infrastructure problem, not a prompt-authoring problem

Any context source without an ingestion and refresh pipeline is degrading infrastructure, so the CLAUDE.md-plus-rules-file practice has a decaying correctness floor and manufactures contradictions the agent cannot detect; the honest unit of comparison is decay rate versus maintenance cost, not effort at authoring time.

Angle

Contrarian argument against the prevailing 'just write a good rules file' advice, framed with freshness and conflict-rate metrics.

Source video

The autonomy tax: how bad context converts into review hours and 1 AM incidents

Each step up the autonomy curve raises the required context quality rather than just swapping tooling, shifting cost from authoring to senior-engineer review and then to incident response; context investment adequate for human-in-the-loop is structurally insufficient for unsupervised agents.

Angle

Staged maturity model with a cost-accounting lens on review tax and silent failure rather than a feature-comparison tour.

Source video

Project ideas

4

ConflictBench

new

A hybrid resolution policy combining source authority with recency weighting yields higher agent answer accuracy on contradictory multi-source context than either recency-only or human-adjudication-last policies.

Proof of concept

Generate synthetic conflicts across code, docs, Slack, and incident records (old architecture record versus last night's Slack thread) and route each through three pluggable resolution policies, scoring agent answers against a ground-truth winning source.

Measurement

Answer accuracy and contradiction-detection rate per policy, plus the fraction of conflicts a policy leaves unresolved or resolves wrongly.

Source video

StopOrGo

beyond-evals

Pre-ranked, engine-side context payloads reduce premature-retrieval-termination rate and total search loops versus an agent driving raw MCP search tools directly.

Proof of concept

Run a fixed task set in two configurations — agent-with-MCP-search versus agent-over-hydrated-engine-payload — instrumenting how many retrieval calls occur before the agent answers and whether the authoritative source was ever read.

Measurement

Premature-termination rate, retrieval-call count per task, and token spend per task in each configuration.

Source video

SilentFailSuite

gatehouse

Agents run without institutional context violate unstated feature-flag and rollout requirements at a materially higher rate than hydrated agents, even when both produce compiling code.

Proof of concept

Build golden tasks whose correct solution depends on a feature-flag or rollout rule that never appears in the prompt, then compare hydrated versus unhydrated agent output against the rule set rather than the compiler.

Measurement

Violation rate per 100 tasks, split by hydrated/unhydrated, with compile-success rate held constant to isolate silent failures.

Source video

ContextRotMeter

movement-lab

Hand-curated markdown context repos decay measurably faster than a maintained ingestion pipeline, but the engine incurs its own staleness and conflict rates that approach curated-repo levels without active monitoring.

Proof of concept

Instrument a curated rules repo and a context engine side by side in an active codebase, logging staleness events and source contradictions over time against a fixed fact set.

Measurement

Staleness half-life in days, contradiction discovery rate per week, and maintenance hours per source per month for each approach.

Source video

Architectural implications

3

Context provenance and governance move out of the repository and into a dedicated ingestion/service layer.

Before

Team context lives in hand-maintained markdown, docs, and tribal knowledge spread across repos and chat.

After

A context engine ingests code, docs, Slack, conversations, and incident management, and serves resolved, permission-filtered context on demand.

Consequence

Context becomes stateful infrastructure with ACLs, freshness guarantees, and schema/versioning concerns rather than prompt-time text.

Source video

Retrieval responsibility shifts from the agent's tool-calling loop into the engine.

Before

Agent decides when to call MCP/vector search and when to stop (leading to satisfaction of search bias).

After

Engine pre-filters, ranks, resolves conflicts, and hydrates a targeted context payload for the agent.

Consequence

Agent design simplifies to reasoning over given context, while reliability depends on engine ranking quality — a different, more testable component boundary.

Source video

Permission enforcement becomes an agent-runtime concern, not just an application concern.

Before

Agents retrieve broadly and rely on the user's own access as an implicit filter.

After

The engine enforces per-user permission and personalized relevance before context is assembled.

Consequence

Context assembly must be user-scoped and auditable, adding identity/ACL coupling to every retrieval path.

Source video

Tradeoffs and failure modes

4

Broad retrieval via MCP tool access

Benefit

Agents get access to information across many systems without building a bespoke ingestion pipeline.

Cost or risk

Satisfaction of search bias leads agents to stop at the first result, producing shallow answers, extra search loops, and token waste; access is mistaken for understanding.

MCP plateau leads to satisfaction of search bias where agents stop at the first piece of information they find.
Open source video
Source video

Hand-curated context repositories

Benefit

Cheap, fast to create, and immediately usable as agent context or rule files.

Cost or risk

They rot faster than they can be updated, producing stale tribal knowledge that conflicts with live sources like recent Slack threads.

Curated-context repos rot faster than you can update them.
Open source video
Source video

Aggressive context hydration and token optimization

Benefit

Large reductions in wall-clock time (2h33m to 25m) and tokens (48% fewer), with better answer quality because signal is not buried.

Cost or risk

Requires standing up and maintaining a multi-source engine with conflict resolution and permission enforcement, which is materially more engineering and governance work than a prompt file.

Using a context engine reduced task time from 2 hours 33 minutes to 25 minutes (2 hours saved).
Open source video
Source video

Increasing agent autonomy (parallel/background agents)

Benefit

Higher throughput and less human-in-the-loop effort per task.

Cost or risk

Bad context converts into review tax on senior engineers and then into silent failures where compilable code violates unstated requirements such as feature-flag rollouts.

An AI agent writing code that compiles but breaks production at 1 AM because it missed feature flag requirements.
Open source video
Source video

Open questions

5

What deterministic mechanism decides which source wins when an old architecture diagram conflicts with a recent Slack conversation with the CTO?

Why unresolved

The summary lists conflict resolution as a requirement and gives the conflict example, but never describes the resolution policy (recency weighting, source authority, or human adjudication).

Research direction

Prototype conflict-resolution policies over synthetic multi-source contradictions and measure agent answer accuracy when the winning source is varied.

Source video

Can a context engine detect silent failures — outputs that compile and look plausible but violate unstated organizational requirements?

Why unresolved

Silent failure is described as a consequence of missing context, but no detection mechanism, eval, or monitoring approach is given.

Research direction

Build golden-task evals over institutional rules (feature flags, rollout policy) to measure how often hydrated vs. unhydrated agents violate requirements that never appear in the prompt.

Source video

Do the reported gains (2h33m to 25m; 20.9M to 10.8M tokens) hold for teams beyond the cited Cloudbeds case, and are they attributable to the engine rather than task selection?

Why unresolved

Only a single case study and two headline metrics are reported, with no baseline methodology, task set, or variance.

Research direction

Replicate the benchmark across multiple repos and task types with a controlled no-engine baseline to isolate the effect of hydration.

Source video

How does 'schema-less lookup' let agents query schemas deterministically, and what does that mechanic actually consist of?

Why unresolved

The claim is asserted as a tool capability without describing the mechanism, index, or query interface.

Research direction

Inspect the open-source Repo Rules Agent implementation and test whether deterministic schema lookup reduces hallucinated field/table references in generated code.

Source video

What is the operational cost of keeping the six engine requirements correct — freshness, ACLs, and conflict policies — as source systems change?

Why unresolved

The engine's maintenance burden is never quantified, even though the curated-context approach is criticized precisely for decay.

Research direction

Instrument a context engine for staleness and conflict rates over time to see whether the engine avoids the same rot it criticizes.

Source video

Key claims

8
comparativeVerification needed

Using a context engine reduced task time from 2 hours 33 minutes to 25 minutes.

Evidence

Using a context engine reduced task time from 2 hours 33 minutes to 25 minutes (2 hours saved).

Question

What task, repository, and baseline conditions produced this measurement, and is it reproducible?

Source video
comparativeVerification needed

Token usage dropped from 20.9 million to 10.8 million tokens (48% fewer).

Evidence

Token usage dropped from 20.9 million to 10.8 million tokens (48% fewer tokens).

Question

Is the 48% figure per-task or aggregate, and does it hold when the same task is run with and without the engine?

Source video
causalVerification needed

The cost of bad context compounds with every step toward autonomy.

Evidence

The cost of bad context compounds with every step toward autonomy.

Question

Can the compounding be measured as increasing review time and incident rate per unit of autonomy?

Source video
causalVerification needed

MCP-based tool access produces satisfaction of search bias, where agents stop at the first piece of information found.

Evidence

MCP plateau leads to satisfaction of search bias where agents stop at the first piece of information they find.

Question

How often do agents terminate retrieval early when an authoritative source exists later in the search space?

Source video
causalVerification needed

Hand-curated context repositories rot faster than they can be updated.

Evidence

Curated-context repos rot faster than you can update them.

Question

What is the observed staleness half-life of curated markdown context repos in active codebases?

Source video
opinionVerification not requested

Access to information does not equal understanding.

Evidence

Access to information does not equal understanding.

Source video
opinionVerification not requested

The gap in agent performance is not intelligence but context.

Evidence

The gap is not intelligence. It's context.

Source video
causalVerification needed

Answer quality improves with a context engine because answers are not buried in noise.

Evidence

Answer quality improves because answers are not buried in noise.

Question

Does removing low-relevance retrieved context measurably increase answer correctness on a fixed eval set?

Source video

Connections

5

context engineMCP

MCP is positioned as a lower plateau; a context engine is the superset that adds conflict resolution, ranking, and permissions the MCP layer lacks.

Source video