Ahead of Main

Disruption Watch

The shifts changing how software gets built and paid for, rated on a four-step ladder and updated every issue. Last updated 26 October 2026.

  1. Repricing the work

    Ultra-cheap and open-weight models

    Meta's Muse Spark Open (816B, 42B active, custom licence) and DeepSeek folding V4-Pro into V4.1 Flash pricing. Record the licence of every open model you recommend, not just its price.

    Evidence log (3)
    1. Muse Spark Open weights; DeepSeek V4.1 Flash absorbs V4-Pro.
    2. Counter-signal: Gemini 4 Argon's introductory price is explicitly temporary.
    3. Haiku 5.5 at $0.10 / $0.50 confirmed on Anthropic's model page, with a reported price step above 100K tokens.
  2. Repricing the work

    Agentic coding and orchestration replacing prompt-level help

    Fork subagents are on by default in scripted Claude Code runs, and 2.1.307 patched a duplicate-work failure mode. The cost effect is unmeasured.

    Evidence log (3)
    1. 2.1.307: each fork told which launch is its own.
    2. Fork subagents on by default for claude -p and the Agent SDK.
    3. Per-subagent compaction and workflow-wide model pinning.
  3. Repricing the work

    Context engineering displacing prompt engineering

    Supporting evidence in every issue so far: idle compaction, per-subagent context windows, cache-aware usage warnings.

    Evidence log (3)
    1. Idle-compaction setting exposed in /config.
    2. Idle compaction and split cache metrics reported.
    3. Per-subagent autoCompactWindow; MCP description cap doubled.
  4. Repricing the work

    MCP and tool protocols standardising the integration layer

    Spec 2026-07-28 is the cloud default. HTTP MCP calls are now retried only for read-only or idempotent tools, so annotations are load-bearing.

    Evidence log (4)
    1. 2.1.310: a timed-out call with an unanswered elicitation now tells the model not to repeat it.
    2. 2.1.307: retries limited to read-only or idempotent tools.
    3. Client fix for servers omitting resultType, ttlMs or cacheScope.
    4. Cloud sessions negotiate spec 2026-07-28 by default; MCP permission-rule matching tightened.
  5. Repricing the work

    App and full-stack generation platforms

    No dated news from the main platforms in the last three issues. Held for lack of evidence, not re-checked.

  6. Hitting the market(new)

    Small decision models as control nodes in agent graphs

    Calibrated yes/no and scoring models are replacing prompted frontier-model calls for route, approve and escalate steps.

    Evidence log (2)
    1. Ollama 0.40.4 runs decision models locally.
    2. New entry. Microsoft-Decision-1 shipped in Foundry on 9 October, after three inference projects added decision-model APIs the week before.
  7. Hitting the market

    Harness and plugin layer: hooks, skills, mods

    Hooks can fail closed, rewrite a subagent at spawn, and demand a human decider. 2.1.309 and 2.1.310 fixed several fail-open cases, so test the failure path of every blocking hook.

    Evidence log (5)
    1. 2.1.309 and 2.1.310 fix hooks that failed open on Windows and in sandboxes. Repair work, no new rung.
    2. 2.1.304: decider "person" and isReadOnly on tool.check.
    3. 12-hook concurrency cap; settable classifier model.
    4. agent.spawn hooks can rewrite model and prompt.
    5. Fail-closed hooks (onFailure: "block").
  8. Hitting the market

    AI-native infrastructure: sandboxes, computer use, cloud dev environments

    Cloud sessions and routines keep maturing; computer-use agents are still rough on Windows.

  9. Hitting the market

    Autonomous QA and test generation

    CheckerBench put a ceiling on it: the best agent wrote a working static-analysis checker for 45% of real CVEs.

    Evidence log (1)
    1. CheckerBench reading: best model and harness 45.33%, C/C++ far behind Python.
  10. Hitting the market(new)

    Portable agent skills (SKILL.md)

    Anthropic donated Agent Skills to the Agentic AI Foundation. Adoption figures are Anthropic's own; security guidance for skill authors and installers is only a roadmap item.

    Evidence log (1)
    1. New entry. Donated to the Agentic AI Foundation on 22 October; Anthropic reports 70+ integrating platforms and about a million published skills.
  11. Accelerating

    Design generation replacing UI and design work

    Standalone Claude Design closes on 14 December 2026. That's consolidation, not a breakthrough.

    Evidence log (1)
    1. Claude Design shutdown date announced in the Claude Code changelog.
  12. Accelerating

    Downward pressure on dev and agency pricing

    Cheaper model tiers support it indirectly. No direct pricing evidence this period.

  13. Watching(new)

    Source control and review built for many agents on one repo

    Cursor Origin is in early beta and GitLab is reported to be reworking its architecture for agent concurrency. No independent benchmarks yet.

    Evidence log (1)
    1. New entry.
  14. Watching(new)

    Observable and auditable agent actions

    Claude Code now emits classifier verdicts and a safety event for each blocked or prompted tool call. One vendor so far.

    Evidence log (1)
    1. New entry. Classifier-verdict fields (2.1.301), decider "person" (2.1.304) and claude_code.safety_event (2.1.307).

How the ladder works

Watching
Early evidence. Worth knowing about, not yet worth acting on.
Accelerating
Multiple independent signals, and the pace is picking up.
Hitting the market
Shipping products you can use or buy today.
Repricing the work
Changing what clients expect to pay for, or what a team needs to staff.

An entry only moves when there's specific evidence, and each issue says what moved it. Entries can be held for lack of new evidence; the issue says when that's the case.