# Agent Update — Crosspost (2026-04-14)

---

## X

Agent update is live.

The main fix: when a tool call failed, the agent used to lose its plan and restart from generic context. Now it keeps the active plan, retries with scoped fallback, and continues from the same state.

Also surfacing blocking constraints earlier in workspace runs — so users stop finding out at step 9.

Changelog: https://doany.ai/changelog/agent-update-2026-04-13

### Adaptation note
Compressed to the single sharpest claim (the retry behavior change). Dropped the bullet lists — they don't land on X. Kept the changelog link. No thread needed; one post covers it.

---

## LinkedIn

Shipped an agent update today. The kind that fixes default behavior, not the kind that needs a narrative.

Problem: when a tool call failed mid-run, the agent dropped its plan and restarted from scratch with generic context. Users saw it as "the assistant forgot what it was doing."

Fix: plan state now survives tool-call exceptions. Retry stays scoped to the failed step — same goal, smaller scope, no full reset. Workspace runs also surface blocking constraints up front instead of late.

In internal smoke runs this week: fewer abandoned runs after API failures, faster completion on messy multi-step tasks.

Full changelog: https://doany.ai/changelog/agent-update-2026-04-13

### Adaptation note
Added the problem/fix framing explicitly since LinkedIn readers include people outside the immediate niche. No closing question, no "lessons learned" wrapper. Kept it operational.

---

## Threads

Agent update shipped.

The thing that kept wasting cycles — tool call fails, agent drops its plan, restarts from zero — that's fixed. It holds the working state now, retries at the failed step, and moves on.

Other changes: workspace runs flag blocking constraints earlier. Cross-platform posting enforces non-duplicate copy by default.

Internal result: fewer abandoned runs, faster completions, less "it forgot what it was doing."

Changelog: https://doany.ai/changelog/agent-update-2026-04-13

### Adaptation note
Conversational reframe of the same core point. Shorter paragraphs, no bullet formatting. Didn't paste the LinkedIn version and trim it — wrote it to read natively on Threads.

---

## Bluesky

Agent update — the retry behavior is fixed.

Previously: tool call fails, agent loses its plan, restarts from generic context. Now: plan state persists through exceptions, retry is scoped to the failed step, working state carries forward.

Also: workspace runs surface blocking constraints up front. Less "why did it only tell me at step 9."

https://doany.ai/changelog/agent-update-2026-04-13

### Adaptation note
Tightest version. "Previously / Now" structure keeps it scannable. No hashtags, no feed-gaming. Link stands alone at the end.

---

## Change Notes (All Platforms)

What shipped:
- Plan state survives tool-call exceptions (previously dropped on failure).
- Retry path is explicit: same goal, scoped to the failed step, no full reset.
- Workspace runs surface blocking constraints early instead of mid-run.
- Cross-platform posting skill enforces non-duplicate copy by default.

What we observed internally:
- Fewer abandoned runs after first external API failure.
- Faster completion on multi-step tasks.
- Reduced "assistant forgot what it was doing" feedback.

---

## Unresolved Publishing Constraints

- UTM links are not generated yet for each platform — the changelog URL above is the bare canonical link.
- Final posting order is undecided (Marketing requested all four today, no time slots set).
- Owner for manual publish is not assigned (Alex vs Marketing coordinator).
