Agentic Coding

COMPLETED April 28, 2026
Summary

Briefing: Agentic Coding

Key Insights

  • The harness, not the model, caused the "Claude regression" you experienced. Anthropic's postmortem on Claude Code quality complaints identified three specific bugs: a caching fix that cleared thinking history on every turn after an idle session (instead of just once), a default reasoning effort downgrade from high to medium, and a verbosity-reduction system prompt that hurt coding quality. If your Claude Code sessions have felt inexplicably dumber since late March, update to v2.1.116+ and verify your effort level is set to xhigh for coding work. This is also the clearest evidence that harness engineering is higher-leverage than model selection — these bugs, not model changes, drove months of user frustration.
  • An update on recent Claude Code quality reports

  • An update on recent Claude Code quality reports

  • Intercom shipped 497 PRs autonomously with a 0.53% revert rate vs. 5.39% for human-authored code. The key was a multi-sub-agent review architecture: separate agents evaluate problem description quality, intent alignment, safety, logical correctness, and best practices independently rather than as a single review pass. The agent also refuses large PRs — flagging anything too big or broadly scoped as a forced quality gate. This isn't theoretical: the architecture is reproducible and the 6–16x faster approval time was measured in their pilot. If you're shipping AI code without multi-agent review, you're reviewing too much manually and catching too little.

  • AI is approving our pull requests: Here's how we made it safe

  • Write custom ESLint plugins instead of agents.md rules for domain constraints. Multiple practitioners independently confirmed that agents ignore markdown rules within minutes of a session. Deterministic linter errors, however, force self-correction every time. Spotify's background coding agent project validated this the hard way: repurposed human documentation failed for context files, but explicit field-mapping tables in structured format produced solid performance across 240 automated migration PRs. The practical rewrite: move your architectural constraints from CLAUDE.md into running validators, and measure the difference. Your effort level is also a context proxy — if you're reaching for max effort on tasks that shouldn't need it, the upstream issue is almost always context quality, not model intelligence.

  • How to Fix Vibe Coding
  • Spec-Driven Dev Is Back. But Not How You Think • Daniel Terhorst-North & Gojko Adzic • GOTO 2026
  • Background Coding Agents: Supercharging Downstream Consumer Dataset Migrations (Honk, Part 4)

  • The Truth About Opus 4.7 Effort Levels

  • GPT-5.5 scored 62.5 on the senior engineer benchmark vs. Opus 4.7's 33 — but only when given an Opus-generated plan. This isn't a clean "GPT-5.5 wins" story. The actual pattern practitioners are converging on: use Opus 4.7 for planning (it writes contract-driven, constraint-heavy specs), then hand the plan to GPT-5.5 for execution (it will actually delete files and start fresh, where Opus patches). GPT-5.5 also trails Opus 4.7 on Terminal Bench (82.7 vs. 69 — but that's GPT ahead), and Opus 4.7 has a new tokenizer that inflates costs 29-47% over what you'd expect. For greenfield product work and design-forward tasks, Opus 4.7 still has better aesthetic sense. For complex refactors and bug-heavy execution, GPT-5.5 is now the better choice — and it's significantly faster, which matters more than people admit for flow state.

  • The Coding Assistant Breakdown: More Tokens Please
  • We Tested GPT-5.5 for 3 Weeks. It's a Beast.

  • Your Prompts Didn't Change. Opus 4.7 Did.

  • AI amplifies what's already there — for better and worse. GitClear found 4.2x more durable code from regular AI users and 84% more code churn and 10x more code duplication across a 211-million-line analysis. These aren't contradictory: teams with good tests, linters, and architectural patterns get AI leverage; teams without them produce more bad code faster. The Anthropic trial showing junior developers who delegated to AI scoring 24-39% on comprehension vs. 65-86% for those who used AI to actively learn is the sharpest version of this. The practical implication: before scaling agent usage, audit your codebase's actual testability and linter configuration — those are the variables that determine whether AI compounds your quality or your debt.

  • Give AI Something Worth Amplifying: Three Priorities for Technical Leaders
  • Full Walkthrough: Workflow for AI Coding — Matt Pocock

Emerging Patterns

  1. CLI-first tools are outgrowing IDE-first tools in actual daily active use at high-adopter companies. Shopify's CTO showed CLI tools growing faster than GitHub Copilot and Cursor, which are not shrinking but not growing at the same rate. At close to 100% DAU across all employees, the growth is entirely in terminal-based agents. The practical advantage: terminal agents run on remote machines, persist when your laptop closes, and can parallelize sessions without IDE overhead. The frontier user pattern is 10-15 parallel sessions managed from tmux rather than one IDE tab. CI/CD is now the stated primary bottleneck — agents can write code faster than existing pipelines can validate it.
  2. CI/CD Breaks at AI Speed: Tangle, Graphite Stacks, Pro-Model PR Review — Mikhail Parakhin, Shopify

  3. The Coding Assistant Breakdown: More Tokens Please

  4. The review crisis is real but solvable with critique loops, not more human eyeballs. LinearB found AI agent-authored PRs wait 5.3x longer for review than human-authored ones. But Intercom's data shows AI-authored PRs, when reviewed by a multi-sub-agent architecture, have lower revert rates than human-reviewed human code. Shopify uses critique loops — one model generates, a different model critiques — and reports significant quality improvement. The teams getting wrecked are rubber-stamping large AI PRs; the teams succeeding enforce small PRs (Intercom's agent refuses large ones), run deterministic validators in their agents' feedback loops, and use cross-model critique before merge.

  5. AI is approving our pull requests: Here's how we made it safe
  6. CI/CD Breaks at AI Speed: Tangle, Graphite Stacks, Pro-Model PR Review — Mikhail Parakhin, Shopify

  7. Give AI Something Worth Amplifying: Three Priorities for Technical Leaders

  8. GitHub's MCP server got a 49% context reduction and 75%+ output token reduction through tool optimization — not better prompts. The lesson generalizes: the most impactful context engineering work is upstream preprocessing and tool surface reduction, not prompt iteration. Cloudflare replaced 7-8 MCP servers with a single TypeScript sandbox, cutting context bloat entirely. Most users don't configure — they use defaults. If you're building on MCP, designing for the default path matters more than building elaborate configuration options nobody will touch.

  9. Lessons from Scaling GitHub's Remote MCP Server — Sam Morrow, GitHub
  10. How to Fix Vibe Coding
  11. MCP = Mega Context Problem - Matt Carey

Dissenting Views

  • On GPT-5.5 as daily driver: practitioners are split on whether it honors intent or just executes literally. The prevailing view from three weeks of testing is that GPT-5.5 is a strong execution engine that follows detailed plans without stopping early. The dissent (direct contradiction, not emphasis difference): several users report GPT-5.5 as "lazy," failing to honor intent, requiring more upfront prompt engineering, and generating context-stuck threads that need frequent kills. SemiAnalysis found it "too conservative" in making code changes — improving token efficiency but hurting correctness. This is likely task-dependent: GPT-5.5 excels when you provide a tight Opus-generated plan; it underperforms when you expect it to infer intent or handle ambiguity.
  • We Tested GPT-5.5 for 3 Weeks. It's a Beast.
  • I don't really like GPT-5.5…

  • The Coding Assistant Breakdown: More Tokens Please

  • On AI code quality improving vs. degrading over time: both are happening, and the determining variable is your review process. The prevailing view from Intercom is that AI code has lower revert rates than human code. The dissent (difference in methodology, not direct contradiction): GitClear's longitudinal data shows 84% more code churn and refactoring dropping from 25% to under 10% of changed lines across 211 million lines of code from 2020-2024. Both are true simultaneously — AI code quality depends entirely on whether you've built the review infrastructure that Intercom built. Teams without multi-agent critique, enforced small PRs, and deterministic validators are accumulating debt, even if individual completions look clean.

  • AI is approving our pull requests: Here's how we made it safe
  • Give AI Something Worth Amplifying: Three Priorities for Technical Leaders
  • AI Inside the Enterprise

Read & Act

What to read

  • The Coding Assistant Breakdown: More Tokens Please — The most rigorous independent Claude Code vs. Codex comparison with actual engineer workflow reports, tokenizer cost analysis showing Opus 4.7's 35% token inflation, and why OpenAI chose "Expert-SWE" over SWE-bench Pro for GPT-5.5's announcement. Read the emerging Codex-for-problem-solving, Claude-for-scaffolding workflow section specifically — that's what SemiAnalysis's team has converged on after GPT-5.5.

  • AI is approving our pull requests: Here's how we made it safe — The architecture details of Intercom's multi-sub-agent review system are the most directly reproducible production case study this cycle. The 0.53% vs. 5.39% revert rate data changes the conversation about whether AI code review can outperform human review — the answer is yes, with the right setup, and the setup is described in enough detail to implement.

  • Full Walkthrough: Workflow for AI Coding — Matt Pocock — The most complete practitioner workflow guide available: the smart zone (~100k tokens) vs. dumb zone model, why compaction is worse than fresh sessions, TDD as essential (not optional) for agents, deep modules vs. shallow modules, vertical slicing, and the "grill me" skill for shared understanding. Dense enough that you'll use it as a reference.

  • An update on recent Claude Code quality reports — If you build on Claude Code, this is essential operational knowledge. Reading the three specific bugs explains months of "forgetful Claude" complaints and tells you exactly what to verify in your setup: effort level, harness version, and whether you're hitting the caching bug in long-idle sessions.

What to do

  • Replace your agents.md domain rules with custom ESLint or linter plugins this week and measure diff. Pick the three rules you most commonly have to re-tell Claude or Codex about — framework-specific patterns, naming conventions, architectural constraints — and write deterministic linter checks for them instead. Run your agent on a feature it would normally struggle with. The change in behavior from getting linter errors vs. ignoring markdown text should be visible within the first session. If you're a TypeScript shop, this is a two-hour investment with compound returns.

  • If you're using Claude Code and experiencing degraded quality, verify your setup against the postmortem checklist before switching tools. Check that you're on v2.1.116+, that your effort level is xhigh for coding work (not medium, which was the broken default), and that you're not hitting the caching bug by starting fresh sessions after long idles rather than resuming them. These fixes resolve issues that drove a meaningful percentage of the "Claude is getting dumber" complaints from March-April.

  • Build a two-agent critique loop for your highest-traffic PR type before scaling agent output. You don't need Intercom's full five-sub-agent architecture to start. Pick your most common agent-generated PR category, write a second agent that checks it against your architectural constraints and test coverage criteria, and run it before merge for two weeks. Measure your manual review time and revert rate. That data will tell you whether to invest further in automated review or whether your current manual process is adequate.

Source Articles

← More from Agentic Coding