Pull to refresh
Logo
What Changed
Following Why Sign Up
AWS gives its coding agents names, memory and a narrow set of permissions

AWS gives its coding agents names, memory and a narrow set of permissions

New Capabilities

Each Kiro bot holds one job and an editable notes file, and reviewers are now reading its default permissions closely

3 days ago: Kam says job scope, not agent IQ, is the measure

Overview

Updated 3 days ago

A developer who starts five coding agents before a morning call comes back to five half-finished conversations. One is stuck waiting to re-approve a tool. Another fixed a test without explaining why. The parallel work happened; the context did not survive it.

Amazon Web Services made Kiro bots generally available on Nov. 5: named agents that own one recurring job, keep an editable notes file and run their own worker sessions inside Kiro Crew. Each starts with only the tools the job needs, and none can grant itself access or hand work to another bot. By Nov. 7, reviewers were reading the defaults closely, noting that new bots start in a permissive Trust mode and that Crew's sandbox leaves the ~/.ssh and ~/.aws folders visible to commands a bot runs.

Why it matters

If persistent memory holds, developers stop re-explaining their codebase every morning and start reviewing finished work instead of supervising sessions.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

1
Memory files per bot
Each bot reads and writes its own markdown notes file, which the developer can open and edit directly.
0
Bots allowed to delegate to other bots
AWS blocks bot-to-bot handoffs so one bot's context does not leak into another's.
4
Example bot roles in AWS's materials
Issue triage, build health, code review and release notes, drawn from the launch post and launch-week coverage rather than the post alone.

Voices

Curated perspectives — historical figures and your fellow readers.

Sign up to generate historical perspectives on this story.

Log in Sign Up

People Involved

KK
Kathy Kam
Director of engineering for agentic AI at Amazon Web Services
Leading the Kiro bots launch and defending its default permissions
 Kyle Seaman
Kyle Seaman
Kiro team, co-author of the launch announcement
Announced general availability of bots in Kiro Crew

Organizations Involved

Amazon Web Services (AWS)
Amazon Web Services (AWS)
Cloud computing provider
Built and shipped Kiro bots

AWS sells cloud infrastructure and developer tools, and now sells persistent agents that run inside its own coding workspace.
KC
Kiro Crew
Developer workspace product
Hosts bots, their dashboards and their worker sessions

The workspace where each bot keeps a chat, a dashboard, a notes file and a set of permitted tools.

Timeline

July 2025 November 2026

5 events Latest: 3 days ago
Tap a bar to jump to that date

  1. Kam says job scope, not agent IQ, is the measure

    Latest Statement

    In an interview with The New Stack, AWS's Kathy Kam said context rot "spreads fast when parallel tasks each hold their own drifting version of reality," and that what an agent is allowed to touch matters more than how capable it is.

  2. Launch-week scrutiny turns to defaults

    Analysis

    Reviewers note bots start in a permissive trust mode and that Crew's default sandbox leaves the ~/.ssh directory visible.

  3. Kiro bots reach general availability

    Product launch

    Named bots get their own notes file, dashboard, schedules and a fixed set of tools; bot-to-bot delegation is blocked.

  4. Kiro Crew arrives with cross-session memory

    Product launch

    A persistent workspace that remembers context between sessions, but still leaves the developer tracking each session.

  5. AWS introduces Kiro

    Product launch

    A spec-driven coding environment where the agent works from written requirements rather than ad hoc prompts.

Scenarios

1

Security researchers demonstrate credential exposure in Kiro Crew's default sandbox

Possible Resolves by Q1 2027

Discussed by: World Programming's launch-week review

The review flags two defaults: bots begin in trust mode, and the sandbox leaves the user's ~/.ssh directory readable. If a researcher turns that into a working demonstration, AWS would have to change defaults or ship a patch, and enterprise buyers would ask for the audit trail first.

2

AWS adds bot-to-bot delegation

Unlikely Resolves by Q3 2027

Discussed by: AWS's Kathy Kam, who said the restriction is deliberate

Kam said bots cannot delegate to each other because handoffs are where context rots. Customer pressure for multi-step pipelines could reverse that. The signal to watch is documentation, not marketing: a changelog entry describing one bot assigning work to another.

3

A major rival ships persistent per-agent memory for coding

Possible Resolves by Q2 2027

Discussed by: Industry coverage framing bots as a competitive move

Kiro Crew's pitch is memory that survives sessions. If that becomes table stakes, GitHub, Google or Anthropic would ship an equivalent: a coding agent with a user-editable memory store scoped to one role. The announcement, not the rumor, is the trigger.

4

AWS publishes a customer running a Kiro bot on a schedule for a month

Uncertain Resolves by Q2 2027

Discussed by: AWS's launch post, which describes scheduled overnight and weekly bot runs

Scheduling is the feature that separates a bot from a chat session. A named customer running one on a recurring schedule for at least 30 days would show the model works outside a demo, and would be the first hard adoption evidence AWS has offered.

5

AWS changes the default permission mode for new Kiro bots

Possible Resolves by Jan 31, 2027

Discussed by: World Programming's review of Kiro's documented defaults

Kiro's own bot-management page says new bots start in Trust, which acts without per-action approval, and Crew's sandbox ships in auto mode with ~/.ssh and ~/.aws visible. If AWS flips that default or narrows the sandbox, the trust critique loses its main evidence. The place to watch is the documentation and changelog, not the launch blog.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

March 2023

AutoGPT (2023)

An open-source project let a language model set its own subtasks and loop through them, with a memory store so it could recall earlier steps. It drew tens of thousands of GitHub stars in weeks.

Then

Demonstrations were impressive and frequently circular; agents repeated work or drifted from the goal.

Now

The pattern of memory plus autonomous looping became standard, but shipped products added narrow scopes and human checkpoints.

Why this matters now

Kiro bots keep the memory and drop the open-ended looping, which is the trade AutoGPT's failures pointed toward.

March 2024

Devin (2024)

Cognition announced an autonomous software engineer that could plan and execute coding tasks end to end, with a demo that set expectations for the category.

Then

Independent evaluations found it solved a minority of real tasks, and early users reported mixed results.

Now

The product persisted with narrower positioning, and the gap between demo and deployment became a cautionary reference for agent launches.

Why this matters now

It is the clearest precedent for the question hanging over Kiro bots: whether a persistent agent's output holds up when a developer is not watching every step.

June 2021

GitHub Copilot (2021)

GitHub released a technical preview of an AI pair programmer that suggested code inside the editor, trained on public repositories.

Then

Adoption spread quickly through individual developers and then teams.

Now

It established the editor as the place where AI assistance lives, and set the expectation that the tool, not the developer, holds the working context.

Why this matters now

Kiro bots extend that same transfer of context one level up, from the file being edited to the recurring job being owned.

Sources

(8)
+2
Kiro VentureBeat The Register World Programming GazeWire Pivot News Dataconomy World Programming