← Latest brief

Brief #222

13 articles analyzed

Practitioners discovered context persistence is now the primary bottleneck, not model capability. They're building shared memory layers, feedback loops, and explicit rationale documentation because intelligence compounds when context survives across sessions—and dies when it doesn't.

Agentic Feedback Loops Compound 180× Production Gains

EXTENDS multi-agent-orchestration — baseline shows coordination patterns, this reveals feedback loops as the mechanism for intelligence compounding

Practitioners achieving production-scale results by feeding agent failures back as context for next iteration, not by upgrading models. 180 merged PRs from daily maintenance loops prove intelligence compounds when context preserves what failed yesterday.

Build explicit feedback capture into agent workflows: log what failed, feed failures as context to next run, measure performance delta across iterations. Don't treat agents as stateless.
A weird experiment I've been trying the last few weeks is having Claude take ...

180 merged PRs across platforms from Claude running daily routines with iterative feedback—'tune its routines so it's better the next day' demonstrates context compounding

Over the last few months I completely rebuilt keep.md from the ground up to s...

Shared persistent memory layer enables agents to 'pick up where we left off' and critique each other's work across sessions

Your sessions have names and can DM each other in Claude Code.

Named sessions with directed messaging preserve context across specialized agents working on shared problem (orders endpoint migration)


Context Decay Solved by Inline Rationale Documentation

EXTENDS context-preservation-across-sessions — baseline shows preservation techniques, this identifies rationale loss as root cause of decay

CLAUDE.md files grow unbounded because deletion requires understanding intent, not mechanism. Practitioners discovered documenting 'why this rule exists' inline prevents context decay and enables confident pruning.

Add inline comments explaining WHY each rule/instruction exists in CLAUDE.md, AGENTS.md, or system prompts. Format: # Rule: [what] | Why: [business reason] | Added: [date] | Owner: [who can answer questions]
Why does CLAUDE.md keep Growing ?

Rules aren't deleted because deletion risk exceeds cost of keeping them—people forget why rules exist. Solution: document rationale inline so next person can confidently modify/remove

Custom Domain Context Depth Defeats Generic Factories

CONTRADICTS agent-orchestration — baseline assumes framework adoption, practitioners are rejecting frameworks for custom context-rich systems

Nine practitioner teams built custom agent orchestration instead of adopting off-the-shelf solutions because generic factories can't capture company-specific intent, culture, or codebase patterns. Context depth, not model capability, determines success.

Stop evaluating agent frameworks by feature count. Instead: document your company's development intent, codebase conventions, and review criteria in structured format, then build custom orchestration that surfaces this context to agents at decision points.
adam is right. 100% right.

Nine custom factories in Swamp discord because aligning intent to AI output requires domain-specific context (company values, codebase structure, development philosophy)

MCP Servers as Local Context Standardization Layer

EXTENDS model-context-protocol — baseline shows MCP as external standard, practitioners are self-hosting for control

Practitioners converting heterogeneous tools to local MCP servers to give agents unified access, proving context engineering infrastructure can be self-hosted and doesn't require cloud dependencies.

Audit which external services your agents repeatedly access. Convert top 5 to local MCP servers on your infrastructure. Use single executor script to route all agent tool requests through MCP interfaces. This compounds: each new agent inherits all integrations automatically.
i just turned the following into MCPs on my mac studio, so all my agents have...

Converted Spotify, Stripe, Anthropic, Twitter, Linear, Notion to local MCP servers with executor.sh routing all agent requests through standardized interfaces

Test-Time Training Enables Within-Session Intelligence Compounding

EXTENDS in-context-learning — baseline shows static few-shot, TTT reveals gradient-based adaptation during inference

Models can improve on novel problems by adapting weights during inference via gradient updates at test time, shifting bottleneck from training data coverage to real-time context utilization.

For repeated tasks where failure patterns emerge, implement explicit test-time feedback loops: capture errors/corrections during inference, inject as context for immediate retry or next iteration. Don't wait for model retraining.
The fact that it works so well for ARC is also a clue to the extent ...

TTT enables models to generalize on unseen problems by updating weights during inference based on test-time context—compounds intelligence within session without retraining

Plan-as-Context Bridges Agent Capability Gaps

New signal

When agents hit tool boundary limitations (UI navigation, non-automatable steps), generating structured plans as explicit context artifacts enables humans or specialized tools to execute, preserving intelligence across the gap.

When agents hit capability boundaries, don't treat it as failure. Generate explicit plan artifacts (structured JSON, markdown checklists) that describe intent and steps, then route to appropriate tool/human. Measure: does plan preserve enough context that execution doesn't require re-explanation?
The most annoying part for me of starting a new project like this is manually...

Created structured plan for Computer Use plugin to automate Google Cloud Console setup steps that fall outside agent terminal capabilities—plan preserves intent across tool boundary

Verification Loop Bottleneck Determines Delegation Success

EXTENDS prompt-engineering — baseline shows prompt structure, this reveals verification as the critical design dimension

Task delegation effectiveness depends on output verifiability, not model capability. Tasks where AI can self-check (visual QA, code execution, self-review) become trivial; judgment-requiring tasks remain human-bottlenecked.

Redesign prompts around verification first: include expected output format, success criteria, self-check questions the model should answer before returning results. Prioritize tasks where verification is mechanical over those requiring taste/judgment.
把 Grok 4.6 当做几周日常主力模型之后,写下自己真正改了什么工作方式

Cursor team member found tasks with easy verification (visual check, code execution) became delegatable; judgment tasks remained bottlenecked—problem isn't capability but verifiability