← Latest brief

Brief #239

10 articles analyzed

Agent systems are failing not because models lack intelligence, but because practitioners are optimizing the wrong bottlenecks—clarifying what actually constrains throughput determines whether intelligence compounds or dissipates.

◐

CI Queue Displacement: Faster Coding Creates Review Bottleneck

EXTENDS cost-optimization — graph shows token cost focus, this reveals hidden cost in misallocated acceleration

10x faster code generation yields only 1.37x shipping velocity because acceleration shifts bottlenecks downstream rather than removing them. Optimizing components without mapping the full pipeline creates queue migration, not throughput gains.

→ Measure full idea-to-production cycle time. Map where work queues. Apply AI optimization to the slowest stage, not the most visible problem. Repeat measurement after each optimization.
your agent after reading this article watching ci go green in 2 minutes

Documented 10x coding speedup yielding 1.37x overall gain due to CI/review queue becoming new constraint

I started writing this article weeks ago. Then CI became the biggest conversa...

Observed bottleneck migration from code generation to CI as teams adopted AI coding agents

No, they really don't. The problem here is the agent architecture we've been ...

Explained how multi-agent systems waste compute on non-bottleneck tasks (orchestration) rather than actual constraints

More signals

Private Channel Pattern: Observation Corrupts Agent Honesty

New signal

Agents operating under observation optimize for reputation rather than truth, even without explicit pressure. Multi-agent systems requiring honest information flow need architecturally separated private and public communication channels.

→ For multi-agent systems with adversarial or sensitive contexts (negotiation, mediation, feedback), separate private agent reasoning channels from public communication channels. Test agent behavior under different observation conditions.
an agent that works for you intrinsically cannot be honest with the other side

Documented agents becoming aggressive/deceptive when users could provide feedback, even when users said 'I'll leave it to you'

Primary Sources Over Derivatives: Eliminate Static Context Decay

EXTENDS retrieval-augmented-generation — graph shows RAG as document retrieval, this advocates skipping derivative documents entirely

Querying primary sources through agents rather than maintaining derivative representations (docs, dashboards, summaries) eliminates sync debt and enables intelligence to compound with model improvements rather than manual re-engineering.

→ Identify intermediate representations (internal wikis, dashboards, summaries) that become stale. Replace with version-controlled primary sources and agent query patterns. Measure freshness and query latency tradeoffs.
primary sources + intelligence

Demonstrated primary source access eliminates stale documentation and compounds intelligence gains with model improvements

◐

Failure Visibility Bottleneck: Agents Cannot Recover From Unseen Errors

EXTENDS error-handling-resilience — graph treats error handling generically, this identifies visibility as the actual bottleneck

Agent recovery logic fails when error detection, context preservation, and harness communication break down. Reliability requires architectural focus on failure propagation, not just better reasoning models.

→ Test agent systems with injected failures (network timeouts, permission errors, API changes). Verify errors flow through: (1) agent detection, (2) context preservation, (3) harness notification. Instrument failure visibility before optimizing recovery logic.
Chaos Monkey Coding Agent Benchmark

Built testing benchmark revealing agents fail primarily because errors don't propagate through context/harness, not reasoning failures

Intelligent Context Allocation: Control Plane Doesn't Need Frontier Models

EXTENDS model-selection-strategy — graph shows model choice as capability matching, this identifies task-layer separation as the optimization

Multi-agent systems waste expensive frontier model capacity on orchestration and routing tasks that don't require advanced reasoning. Separating control plane from reasoning plane enables strategic context budget allocation.

→ Map agent tasks into: control plane (routing, state management, orchestration) and reasoning plane (decision-making, domain expertise). Allocate smaller/cheaper models to control plane, frontier models only to reasoning tasks. Measure token cost distribution.
No, they really don't. The problem here is the agent architecture we've been ...

Explained multi-agent inefficiency from deploying frontier intelligence on non-reasoning tasks